> 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/2026/10.md).

# October 2026

MaiAgent platform October 2026 release notes

## v2026.10.19 <a href="#v2026-10-19" id="v2026-10-19"></a>

***

### 📌 Highlights

* **\[Shared] Direct OpenAI GPT-5.6 Sol stays available**: The directly connected OpenAI GPT-5.6 Sol model is no longer disabled by the post-deployment schedule, so AI assistants using it are not automatically moved back to the Azure endpoint.
* **\[Shared] Word-by-word streaming for OpenAI models on Bedrock**: Answers from AI assistants using GPT-6.1 Sol and GPT-5.6 Sol on Bedrock now appear word by word instead of all at once after the full response is complete.
* **\[Build] Daily and monthly Credits spending limits for API keys**: Admins can set a daily and a monthly Credits spending limit for each API key, and calls are rejected once a limit is exceeded. [User Guide](https://docs.maiagent.ai/build/ai-gateway)
* **\[Serve] Internal conversation notes**: Teammates can leave notes in a conversation that only internal members can see, synced in real time to other online teammates. [User Guide](https://docs.maiagent.ai/serve/daily-operations/reply)
* **\[Shared] Vision Parser economy tier**: Knowledge base files can use the Vision Parser economy tier, billed per page, with the per-page price shown in the parser list. [User Guide](https://docs.maiagent.ai/build/knowledge-base)
* **\[Build] Assistant tool loading mode: automatic, all, or dynamic search**: AI assistants can be set to load tools automatically, load all of them, or search for them dynamically, and tools found by search are kept for the whole conversation. [User Guide](https://docs.maiagent.ai/build/agent-builder)
* **\[Serve] Automatic conversation splitting and follow-up summaries for integrated messaging platforms**: Conversations on messaging platforms such as LINE and Messenger are split automatically after being resolved or going idle, and the new conversation carries a follow-up summary. [User Guide](https://docs.maiagent.ai/serve/daily-operations/reply)
* **\[Work] Dedicated panel for MaiGPT document translation**: MaiGPT now presents document translations in a dedicated translation card, translating paragraph by paragraph and showing the source file details along with the source and target languages. [User Guide](https://docs.maiagent.ai/work/maigpt)
* **\[Work] Schedules auto-disable after consecutive failures, and short intervals require a stop condition**: A schedule is disabled automatically, and its creator notified, when consecutive failures reach the limit or the end date arrives. Schedules that run more often than once an hour must have a stop condition. [User Guide](https://docs.maiagent.ai/work/overview)
* **\[Work] AI assistant file changes run only after user confirmation**: In organizations that have not enabled skipping confirmation, MaiGPT first lists its planned file changes and only creates folders, renames, and moves files after the user confirms.

### 🗂️ Full Updates (by Product Line)

<details>

<summary>Serve (51 items)</summary>

#### ✨ New Features

* **\[Serve] Internal conversation notes**: Teammates can leave notes in a conversation that only internal members can see, synced in real time to other online teammates. [User Guide](https://docs.maiagent.ai/serve/daily-operations/reply)
* **\[Serve] Automatic conversation splitting and follow-up summaries for integrated messaging platforms**: Conversations on messaging platforms such as LINE and Messenger are split automatically after being resolved or going idle, and the new conversation carries a follow-up summary. [User Guide](https://docs.maiagent.ai/serve/daily-operations/reply)
* **\[Serve] Assistant sign-in source supports organization login settings and AD name validation**: Messaging platform login settings gain a new MaiAgent organization login settings source, and the SSO name for AD is now validated when saved. [User Guide](https://docs.maiagent.ai/serve/daily-operations/reply)
* **\[Serve] Auto-close idle live-agent conversations**: Messaging platforms can enable idle auto-close, so a conversation closes automatically when the customer does not reply within the configured number of minutes after an agent response. [User Guide](https://docs.maiagent.ai/build/conversation-settings/human-handoff)
* **\[Serve] Messenger and Instagram ice breakers and persistent menu**: Messenger and Instagram messaging platforms can set ice-breaker questions per language and sync the assistant's quick menu as a persistent menu. [User Guide](https://docs.maiagent.ai/build/choose-channel)
* **\[Serve] Voice support supports gpt-realtime-whisper**: Voice assistants gain OpenAI gpt-realtime-whisper as a streaming speech recognition option. [User Guide](https://docs.maiagent.ai/serve/voice)
* **\[Serve] Messaging platform Access Tokens can be replaced**: People with permission to create messaging platforms can replace the credentials of LINE, Telegram, Viber, Slack, and Microsoft Teams messaging platforms directly, without deleting and recreating them. [User Guide](https://docs.maiagent.ai/serve/daily-operations/reply)
* **\[Serve] Location request timing setting for web chat**: Messaging platforms gain a location request timing setting, which can be changed so the visitor's location is requested only when the AI asks through a question card. [User Guide](https://docs.maiagent.ai/serve/daily-operations/reply)
* **\[Serve] Conversation API accepts an external user ID directly**: When sending a message through the conversation API, an external user ID can be supplied directly, and the system automatically finds or creates the matching contact. [User Guide](https://docs.maiagent.ai/org/developer)
* **\[Serve] LINE replies can include a sticker**: When enabled on a LINE messaging platform, an AI reply can include one LINE sticker along with its text. [User Guide](https://docs.maiagent.ai/build/choose-channel)
* **\[Serve] Assistant cards on Messenger and Instagram are sent as native cards**: Card messages from the assistant are now shown as native Meta cards on Messenger and Instagram. [User Guide](https://docs.maiagent.ai/build/choose-channel)
* **\[Serve] Internal conversation notes**: Conversations gain internal notes that agents can switch to in the timeline and input bar, and view in the contact sidebar. [User Guide](https://docs.maiagent.ai/build/conversation-settings/human-handoff)
* **\[Serve] Messaging platform Access Tokens can be changed**: The edit page of messaging platforms such as LINE, Telegram, Viber, and Microsoft Teams lets you change the Access Token or App Password directly. [User Guide](https://docs.maiagent.ai/build/choose-channel)
* **\[Serve] Automatic conversation splitting for integrated messaging platforms**: Integrated messaging platforms gain an auto-split settings tab, and the conversation list marks new and old conversations and shows the follow-up summary. [User Guide](https://docs.maiagent.ai/serve/daily-operations/reply)
* **\[Serve] Messenger/IG ice breakers and persistent menu**: The Messenger and Instagram settings pages can configure ice-breaker questions and a persistent menu, and sync them to the platform. [User Guide](https://docs.maiagent.ai/build/choose-channel)
* **\[Serve] Idle auto-close for messaging platforms**: The dispatch settings of a messaging platform can enable idle auto-close with a custom idle duration and closing message, and the timeline shows the close record. [User Guide](https://docs.maiagent.ai/serve/daily-operations/reply)
* **\[Serve] Location requests on question cards**: Location awareness for messaging platforms gains a location request timing setting, so location can be captured when a customer answers a question card. [User Guide](https://docs.maiagent.ai/serve/daily-operations/reply)
* **\[Serve] Satisfaction surveys for internal messaging platforms**: Internal messaging platforms support satisfaction surveys, so users can give feedback on their satisfaction after a conversation. [User Guide](https://docs.maiagent.ai/serve/daily-operations/reply)
* **\[Serve] Notice when AI auto-reply is paused for lack of credits**: When AI auto-reply pauses because credits ran out, the system shows a global banner and a message label, and the notice text can be customized. [User Guide](https://docs.maiagent.ai/others/credits)

#### 🚀 Core Performance Improvements

* **\[Serve] Faster inbox unread count loading**: When non-owner members open a large inbox such as MaiGPT, unread counts now use an indexed query, so they no longer occasionally wait over ten seconds.
* **\[Serve] Faster loading of the all-conversations list**: The all-conversations list now uses an index that holds only conversations with incoming messages, so it no longer times out wading through many empty conversations.
* **\[Serve] Faster voice call setup**: Voice support now obtains the call connection from a pre-built room, shortening the wait.

#### 😊 User Experience and Interface Improvements

* **\[Serve] Conversation analysis summary written automatically on handoff to a human**: When a conversation is handed to a human and the platform has conversation analysis enabled, the details automatically show the AI summary.
* **\[Serve] Platform API error responses now carry fixed error codes**: When the send-message API hits a contact error, it returns a fixed error code so callers can tell the cause.
* **\[Serve] Audio and video attachments playable in admin conversation records**: Audio and video attachments in the admin's conversation records can be played directly.
* **\[Serve] Enable OCR toggle removed from messaging platforms**: The retired Enable OCR toggle is removed from messaging platform forms, and the system now detects how to parse attachments automatically.
* **\[Serve] Greeting messages no longer show like/dislike**: Greeting messages in a conversation no longer show like and dislike buttons.

#### 🔐 Security and Governance Enhancements

* **\[Serve] Webhook and Slack credentials stored encrypted**: The Secret Token of message webhooks and the Bot Token and Signing Secret of Slack are now stored encrypted.
* **\[Serve] Handling of disabled Web Chat IDs at the login entry**: A Web Chat ID disabled by another organization now gives the same result at the login entry as an ID that does not exist.
* **\[Serve] Telegram channel verifies the push source**: The Telegram channel now verifies where pushes come from, and forged push requests are rejected.

#### 🛠️ Bug fixes

* **\[Serve] File transfers on messaging platforms download in the background and report failures**: When end users send files on messaging platforms, the files are now downloaded in the background with a size limit, and users are notified if a download fails, so other messages are no longer lost along with it.
* **\[Serve] Failed Web Chat message submissions no longer leave incomplete data**: When Web Chat sends a message, the related data is committed together, and if any one write fails the whole set is rolled back, so no incomplete state is left behind.
* **\[Serve] Image URLs without a file extension are now sent as images**: GitBook screenshots referenced in replies now appear as images on LINE, Messenger, Instagram, WhatsApp, and Team+.
* **\[Serve] Telegram typing indicator and LINE WORKS bot rename handling**: In Telegram groups the typing indicator now appears consistently when certain documents are received, and a LINE WORKS bot that has been renamed is recognized automatically under its new name.
* **\[Serve] LINE sticker tags no longer leak**: Sticker tags in AI assistant replies are no longer shown as raw text on any conversation channel.
* **\[Serve] Voice and video from channels can be played directly**: Voice and video received from channels can now be played directly in the conversation.
* **\[Serve] Fixed LINE emoji placeholder text**: Emoji placeholder text in LINE messages is no longer mistaken by the AI for a person's name.
* **\[Serve] Deep Research keeps original file names when analyzing photos**: During Deep Research, the internal assistant now sees the original file name of every photo uploaded by a customer, for example "receipt.png".
* **\[Serve] Gemini now resends automatically after an abnormal finish, so LINE no longer occasionally goes unanswered**: When Gemini ends with an unrecognized finish code and an empty reply, the request is resent automatically, so LINE conversations no longer occasionally get no response.
* **\[Serve] Abnormal attachment IDs are no longer silently dropped when sending a message**: When a message is sent with a missing or malformed attachment ID, the attachment is no longer silently discarded and lost.
* **\[Serve] Full-width ＠ mentions in groups can also trigger the assistant**: Mentioning the assistant with a full-width ＠ followed by its name in LINE, LINE WORKS, and Telegram groups is now treated as a valid mention.
* **\[Serve] Replies on Telegram and other channels no longer include code and raw data**: AI replies on Telegram and other channels no longer include the code and raw data the assistant wrote into files.
* **\[Serve] AI replies are no longer interrupted by background machine scale-down**: AI replies on Telegram, LINE, and other channels are no longer cut off when background machines scale down, so users receive their replies normally.
* **\[Serve] The reset command on every channel only opens a new conversation**: Entering the reset command on any conversation channel now only opens a new conversation; it is no longer passed to the AI for a reply or charged against your quota.
* **\[Serve] Fixed blank conversation record panel for LINE, Slack, and Teams**: Conversations on the LINE, Slack, and Teams channels now display their content correctly in the back-office "Full conversation record" panel.
* **\[Serve] Fixed Telegram receiving file-write content under rewriting Hooks and masking**: When an organization has a rewriting Hook or content masking enabled, Telegram and other channels no longer receive file-write content, and memory now stores file pointers instead.
* **\[Serve] Email channel retries daily after a login rejection**: After a login is rejected, the Email channel now retries automatically every day instead of staying stopped permanently.
* **\[Serve] Claude 5.5 can be required to include text with Flex cards**: When "Require text with cards" is enabled for Claude 5.5 Sonnet and Opus, the request is retried to obtain reply text, so a card is no longer sent alone.
* **\[Serve] Fixed a lone "\[" on its own line disappearing from AI replies**: A "\[" on a line by itself in an AI reply no longer disappears in the chat window.
* **\[Serve] Fixed messages disappearing when the last line contains only digits**: In Web Chat and conversation records, messages whose last line contains only digits (such as a phone number) no longer disappear.
* **\[Serve] Fixed suggested questions not updating after switching language**: In Web Chat, suggested questions now switch to the new language after the language is changed.

</details>

<details>

<summary>Build (83 items)</summary>

#### ✨ New features

* **\[Build] Daily and monthly Credits spending limits for API keys**: Administrators can set daily and monthly Credits spending limits for each API key, and calls are rejected once a limit is exceeded. [User guide](https://docs.maiagent.ai/build/ai-gateway)
* **\[Build] Assistant tool loading can be set to automatic, all, or dynamic search**: In the settings, an AI assistant can choose to load tools automatically, load all of them, or search for them dynamically, and tools found by search are kept for the whole conversation. [User guide](https://docs.maiagent.ai/build/agent-builder)
* **\[Build] AI assistants can specify their own question rewriting model**: In an AI assistant's advanced settings you can choose a question rewriting model; if none is selected, the system setting is used. [User guide](https://docs.maiagent.ai/build/agent-builder)
* **\[Build] Files returned by MCP tools are saved automatically as conversation attachments**: Images, audio, and documents returned by MCP tools are saved automatically as conversation attachments that the AI can read directly. [User guide](https://docs.maiagent.ai/build/tools)
* **\[Build] Agent health check**: You can run a health check on an AI assistant, which lists problems in its instructions, skills, and model settings along with suggested fixes. [User guide](https://docs.maiagent.ai/build/agent-builder)
* **\[Build] Files downloaded by tools are saved automatically as conversation attachments**: Files returned by API tools or MCP tools (such as PDFs, Excel files, and images) are saved automatically as conversation attachments, and the assistant can read their content directly. [User guide](https://docs.maiagent.ai/build/tools)
* **\[Build] Duplicate file detection for knowledge base uploads**: When a file with identical content is uploaded to the same knowledge base, the system rejects the duplicate upload. [User guide](https://docs.maiagent.ai/build/knowledge-base)
* **\[Build] OCR on scanned pages enabled by default, and per-page billing for MaiAgent Parser scanned pages**: Document parsing now enables the OCR reference for scanned pages by default, and MaiAgent Parser also adds OCR for scanned pages and bills per page. [User Guide](https://docs.maiagent.ai/build/knowledge-base)
* **\[Build] MaiAgent Parser supports HTML files**: MaiAgent Parser can now parse .html and .htm files, for both knowledge base uploads and conversation attachments. [User Guide](https://docs.maiagent.ai/build/knowledge-base)
* **\[Build] Auto-detect now classifies page structure page by page**: Auto-detect for knowledge base uploads now classifies each page, scanned files go straight to Vision, and shared images and logos are no longer misjudged. [User Guide](https://docs.maiagent.ai/build/knowledge-base)
* **\[Build] AI assistant defaults**: Organizations can set AI assistant defaults that are applied automatically to new assistants and can be enforced on existing assistants. [User Guide](https://docs.maiagent.ai/build/agent-builder)
* **\[Build] Parser options for on-premises trial: vLLM and dots.mocr**: Knowledge base parsing can use an OpenAI-compatible (vLLM) endpoint and dots.mocr to recognize scanned pages, and the system status page also shows the three endpoints, with a dormant endpoint not counted as a failure. [User Guide](https://docs.maiagent.ai/build/knowledge-base)
* **\[Build] Parsers for on-premises trial**: When uploading files to a knowledge base, you can choose OCR (GPU) and two Vision parsers, Gemma 4 and Qwen 3.8, to try the parsing results of the on-premises plan first. [User Guide](https://docs.maiagent.ai/build/knowledge-base)
* **\[Build] Agent health check batch ranking and one-click fixes**: Administrators can run a health check on all editable AI assistants at once, see results ranked by severity, and apply the suggested fixes. [User Guide](https://docs.maiagent.ai/build/agent-builder)
* **\[Build] Question rewriting model can be set separately**: The advanced settings of an AI assistant now include a Question Rewriting Model dropdown, which follows the system setting by default. [User Guide](https://docs.maiagent.ai/build/agent-builder)

#### 🚀 Core Performance Improvements

* **\[Build] Faster code sandbox execution**: The code execution tool reuses the sandbox within a single reply turn instead of starting and stopping it on every call.
* **\[Build] Skill resource package cache**: Skills with resource files share a cache by version, so each conversation no longer downloads and packages them again.
* **\[Build] MaiAgent Parser (Offline) runs OCR only on blocks that need it**: When MaiAgent Parser (Offline) parses digital PDFs, tables and shaded blocks that already contain text are read directly instead of being recognized again.
* **\[Build] Browser panel frames sent only to people watching**: The Browser Tool's browser panel frames are sent only to the users viewing them, limited to at most 5 frames per second.
* **\[Build] MCP tools share one connection per conversation turn**: When the same MCP server is called several times in one conversation turn, a single connection is shared, reducing the wait caused by repeated handshakes.
* **\[Build] Better prompt cache hits for Claude assistants with many tool calls**: For assistants using Claude in conversations with many tool turns, the next turn can reuse the cache up to the previous question instead of rewriting the entire history.
* **\[Build] Standard replies on OpenAI/Azure exclude knowledge base retrieval content from the cache prefix**: For standard replies on OpenAI and Azure, knowledge base retrieval content is moved out of the cache prefix, which improves prompt cache hits.
* **\[Build] Test dataset evaluation saves re-query costs**: When test dataset evaluation receives a JSON reply in text format, it uses it directly, cutting Sonnet 5.5 evaluation costs by about 37%.
* **\[Build] Hook sandbox warm-up**: Platform administrators can turn on Hook sandbox warm-up in system settings to shorten the wait for each Hook check (off by default).
* **\[Build] Faster Credit totals on the AgentOps text assistant overview**: The Credit total on the AgentOps text assistant overview now queries only the organization's own transactions, so large organizations can see the value too.
* **\[Build] AI Gateway connection test reports upstream overload quickly**: When an AI Gateway custom model connection test hits an upstream overload, the cause is reported in about 1 second instead of after a two-minute wait.
* **\[Build] Short PDFs no longer queue in Vision parsing**: When several short PDFs are uploaded at once with the Vision Parser, they can be parsed at the same time instead of queuing because the memory estimate was too high.
* **\[Build] Faster content safety Hook and human handoff decisions**: The content safety Hook now sends multiple pieces of content in one request, and Hook decisions and human handoff decisions explicitly turn off thinking for faster responses.

#### 😊 User Experience and Interface Improvements

* **\[Build] AgentOps can list only organization-created assistants**: The AgentOps assistant list can show only the assistants created by the organization and hide the official built-in assistants.
* **\[Build] Tool search now keeps only long-tail tools**: Assistants with tool search enabled now get platform features and common tools directly, and only long-tail tools need to be searched for.
* **\[Build] Knowledge base retrieval tool supports multiple queries at once**: The knowledge base retrieval tool can handle several queries at once and groups the results by query.
* **\[Build] Vision parsing attaches an OCR reference for scanned pages**: For scanned pages, the SaaS Vision parser first obtains Azure OCR results for the model to cross-check, improving recognition accuracy for proper strings.
* **\[Build] The reason OCR was skipped on scanned pages is now visible**: Knowledge base files show why OCR was not run on scanned pages, such as insufficient credits or an OCR service that is not configured.
* **\[Build] Skill reference file reads are no longer truncated**: When an assistant reads a skill reference file, the result is no longer truncated and the full content is given to the model.
* **\[Build] Knowledge base files show the decision record of the auto-detected parser**: Knowledge base file information now shows the parser selected by auto-detect and the reason, and records any change the user makes afterward.
* **\[Build] Hook execution logs show external call duration**: Hook execution logs and exported files now list the name, duration, and success or failure of each external call.
* **\[Build] Parser menu cleanup**: The parser menu in file uploads has a new order, MaiAgent Vision Parser is renamed, the CPU version is listed last, and retired options are removed.
* **\[Build] Parser menu shows pricing basis and reference prices**: The parser menu adds an explanation of the pricing basis, and the Vision and on-premises trial parsers show a reference price while no measured data is available yet.
* **\[Build] Agent Marketplace layout redesign**: The Agent Marketplace moves to a new layout with a title bar, a category sidebar, and shared search, with cards, recently used items, and the preview entry point updated alongside, and the mobile layout adjusted to a 4-column breakpoint.
* **\[Build] AI assistant list adds a creator filter**: The AI assistant list gains a "Creator" filter, and the name column now wraps instead of being truncated.
* **\[Build] Knowledge base files show an auto-detect badge**: The knowledge base file list shows an "Auto" badge and the reason the parsing method was selected automatically.
* **\[Build] AI assistant analytics shows satisfaction as a percentage**: Satisfaction in AI assistant analytics is now displayed as a percentage on a 0–100% axis.

#### 🔐 Security and Governance Enhancements

* **\[Build] Calls are rejected when a self-hosted model's credential is disabled**: When the credential of a self-hosted model is disabled or missing, calls are rejected instead of falling back to the platform key.
* **\[Build] URL parameters of MCP tools are also checked for attachment ownership**: If a URL-type parameter of an MCP tool points to an attachment from another conversation or organization, that call is blocked.
* **\[Build] MCP attachment URL path-rewriting bypass closed**: The attachment ownership check for MCP tools now blocks URLs that use path rewriting to point to attachments in other people's conversations.
* **\[Build] Bring-your-own vector database keys and passwords are stored encrypted**: The API key of a bring-your-own vector database and the connection passwords for OpenSearch, Oracle, and SynxDB are now stored in encrypted form.
* **\[Build] High-cost models must be explicitly checked on the key**: For designated organizations, an API key must explicitly check a high-cost model such as Opus in its available models before it can call that model.
* **\[Build] Assistant database connection settings are stored encrypted**: The database connection settings of an assistant are now stored in encrypted form.

#### 🛠️ Bug Fixes

* **\[Build] Agent replies cut off midway can be resumed**: After an Agent reply is cut off by a timeout, the user can say "continue" to pick up from the earlier progress.
* **\[Build] Blank closing after delegation fixed**: In an AI assistant team, even if the source assistant ends with a blank closing after delegating a task, the answer from the delegation target is relayed.
* **\[Build] AgentOps viewing a colleague's MaiGPT conversation fixed**: An AgentOps monitor who opens a colleague's MaiGPT conversation from the record details no longer gets a page-not-found error.
* **\[Build] Progress is summarized when the tool call limit is reached**: When the tool call limit is used up and the model leaves no text, the model is asked once more to report its current progress.
* **\[Build] Vision parsing attaches only the images actually shown on each page**: When the Vision parser handles a PDF exported from a presentation, each page now refers only to the images drawn on that page, and no longer mixes in images from other pages.
* **\[Build] Gemini no longer fails to reply when a tool name starts with a digit**: When an API tool name starts with a digit, assistants that use Gemini no longer fail the entire reply.
* **\[Build] Prompt injection protection no longer blocks skill expansion and tool search**: After an organization enables prompt injection protection for Agent assistants, the assistant can still expand the organization's own skills and search for tools normally.
* **\[Build] The web fetch tool validates multi-passage quotes before returning the full page**: The web fetch tool now validates multi-passage quotes first, and only when validation fails does it return the full page content to the model.
* **\[Build] The closing message is no longer blank when the tool use limit is reached**: When an assistant uses up its tool calls for the turn and the model leaves no visible content, a default closing explanation is shown instead.
* **\[Build] Refreshing the tool list while editing an MCP tool no longer returns 401**: When editing an MCP tool, clicking "Refresh" without modifying the headers no longer produces an HTTP 401 error.
* **\[Build] Conversation attachments using the Offline parser are now parsed by a worker**: When an attachment uploaded in a conversation uses MaiAgent Parser (Offline), it is actually parsed by that parser instead of being silently swapped for another one.
* **\[Build] File retrieval and concurrent file reads no longer interfere with each other**: When the AI runs a retrieval limited to specific files while reading other files, the reads are no longer wrongly narrowed in scope and no longer fail.
* **\[Build] A failed tool search now points to skills or existing tools**: When a tool search finds nothing, the AI is prompted to expand a skill or call a tool it already has directly, and no longer wrongly replies that the feature does not exist.
* **\[Build] Conversations created by an API Key can use MCP tools under the caller's identity**: Conversations invoked with an API Key can now use MCP tools as the member who owns that Key, instead of always failing authorization.
* **\[Build] Pages blocked by Vision parsing content filters are filled in from the text layer**: When the Vision Parser meets a page blocked by a content filter, it fills the page in with text-layer content, so the page no longer disappears entirely.
* **\[Build] Team delegation relay pauses only when the user sees the file panel**: The relay of a team delegation pauses only when the user actually sees the file panel, and otherwise the full reply is kept.
* **\[Build] Transient Azure tool-argument streaming errors are retried automatically**: Assistants using Azure models now automatically resend when a transient streaming error occurs during a long tool workflow, instead of stopping.
* **\[Build] Skill instructions stay intact under size-based truncation**: For assistants with size-based truncation enabled, skill instructions are kept whole after a skill is expanded and are no longer cut off.
* **\[Build] Evaluation custom metrics defer generation when steps are empty**: When the judge model returns empty steps for a custom metric in an evaluation dataset, generation is now deferred, and the whole batch is no longer sent back.
* **\[Build] Fixes for message search and segmented reading of long messages**: Message search no longer fails because of regular expression errors, and long messages over 1,000 characters can now be read in full, segment by segment.
* **\[Build] Fixes for OCR detection and billing in document parsing**: Document parsing now runs OCR only on genuinely scanned pages and bills for it correctly, and presentation files no longer trigger an extra model call for the decision.
* **\[Build] Database connection test now times out within 10 seconds**: When a new database connection points to an unreachable host, the connection test shows "Connection failed" within 10 seconds instead of hanging for about two minutes.
* **\[Build] AI Assistant Monitoring text assistant overview no longer times out for large organizations**: The AgentOps text assistant overview now loads for large organizations too, and shows "-" instead of an error when Credit consumption cannot be calculated in time.
* **\[Build] Fixed role permissions being lost when duplicating an AI assistant**: Duplicating an AI assistant now keeps each role's original read-only, edit, and delete permissions.
* **\[Build] OCR of Traditional Chinese scans no longer mixes in Simplified characters**: After parsing with Docling and MaiAgent Parser, OCR results for Traditional Chinese scanned documents are now consistently in Traditional characters.
* **\[Build] Unresponsive external MCP servers now wait at most 30 seconds**: When an external MCP server does not respond, the assistant now replies after at most 30 seconds instead of waiting 5 minutes.
* **\[Build] Replies intercepted by a Hook are no longer misclassified as errors**: When a Hook intercepts an AI assistant's reply card, the conversation and evaluation results are now marked as blocked instead of being reported as a system error.
* **\[Build] Fixed tool result attribution after an Agent Teams handoff**: After an Agent Teams handoff, the tool call results of the receiving assistant are no longer wrongly replaced by the previous assistant's output.
* **\[Build] Testing the on-premises parser on the system status page no longer wakes a sleeping GPU**: When an administrator tests the trial parser on the system status page, a sleeping endpoint now shows "Sleeping" and is no longer woken up or alerted on.
* **\[Build] Longer cold-start wait for the trial parser**: The cold-start wait limit for the trial parser is extended to 10 minutes, and the description now reflects the measured three to six minutes.
* **\[Build] Earlier masking is preserved when a Hook blocks or skips input**: When a Hook blocks or skips an input, the masking already applied by a previous Hook is kept, so sensitive content is not exposed.
* **\[Build] OCR parser straightens skewed scanned pages before reading tables**: The OCR parser now straightens skewed scanned pages before reading tables, so long tables no longer end up with misaligned rows.
* **\[Build] Manual-selection-only parsers are skipped when Vision parsing falls back**: When Vision parsing fails and falls back to MaiAgent Parser, parsers that users must pick themselves are no longer selected.
* **\[Build] Fixed multi-turn evaluation cards not updating with progress**: While a multi-turn conversation evaluation is running, the conversation cards now refresh as it progresses.
* **\[Build] Fixed AI Assistant Monitoring log row settings not being retained**: Row settings for the text and voice conversation logs in AI Assistant Monitoring are now kept after a refresh and after logging in again, and the two tables no longer overwrite each other.

</details>

<details>

<summary>Shared (78 items)</summary>

#### ✨ New Features

* **\[Shared] Vision Parser Economy tier**: Knowledge base files can use the Vision Parser Economy tier, which is billed per page, and the parser list shows the price per page. [User Guide](https://docs.maiagent.ai/build/knowledge-base)
* **\[Shared] Role permission governance opened to all organizations**: Role permission governance is now available to every organization, with bulk permission settings on the role page and effective permissions viewable on the member page. [User Guide](https://docs.maiagent.ai/org/roles)
* **\[Shared] Added GLM-5.3 and preserved Azure thinking content**: GLM-5.3 is added to the model menu, and thinking models on Azure can now show their thinking content in conversations and carry it into the next turn. [User Guide](https://docs.maiagent.ai/build/ai-gateway)
* **\[Shared] Added Cohere Embedding and Reranker on Azure and Bedrock**: Embedding and Reranker models now offer Cohere options from the Azure and Bedrock providers, with the same credit deduction as the models of the same name. [User Guide](https://docs.maiagent.ai/build/ai-gateway)
* **\[Shared] Qwen3.8 Max enabled**: Qwen3.8 Max is now selectable in assistant and model access permissions, succeeding the retired Qwen3.6 Max. [User Guide](https://docs.maiagent.ai/build/ai-gateway)
* **\[Shared] Embedding and Rerank included in organization model access permissions**: Organization owners can switch Embedding and Rerank models on or off in model access permissions, and disabled models no longer appear in the knowledge base settings menu. [User Guide](https://docs.maiagent.ai/org/overview)
* **\[Shared] Qwen3.6 27B enabled**: Qwen3.6 27B is now selectable in the assistant model menu, the same model that ships on-premises. [User Guide](https://docs.maiagent.ai/build/ai-gateway)
* **\[Shared] Organization owners can rename members within the organization**: Organization owners can change a member's name within the organization through member editing, the API, or Excel import, and every rename is recorded in the activity log. [User Guide](https://docs.maiagent.ai/org/overview)
* **\[Shared] Back-office system status page (external service checks)**: The back office has a new system status page that checks external services daily and flags the features degraded as a result. [User Guide](https://docs.maiagent.ai/org/overview)
* **\[Shared] Conversation attachments automatically pick a parser by content**: Conversation attachments now automatically choose a parser based on their content, and the "Enable OCR" option on the chat platform is removed. [User Guide](https://docs.maiagent.ai/build/knowledge-base)
* **\[Shared] Organization-level AI assistant defaults**: Organizations get settings storage and a dedicated permission for AI assistant defaults, so administrators can read and write organization-level default settings. [User Guide](https://docs.maiagent.ai/org/overview)
* **\[Shared] Organization announcements can remind on every login**: Organization announcements support a "every login during the announcement period" reminder frequency; members can choose to stop being reminded, and owners can remind everyone again. [User Guide](https://docs.maiagent.ai/org/overview)
* **\[Shared] Model capability scores and removal of input processing speed**: The model access permissions page now shows a capability score for each model, and the inaccurate input processing speed is removed. [User Guide](https://docs.maiagent.ai/build/ai-gateway)
* **\[Shared] Successor model automatically added to API key allowlists on model retirement, with notifications**: When a model is retired, the system automatically adds its successor model to the AI Gateway allowlist of affected API keys, and notifies the key holder, the organization owner, and customer support if the successor cannot be added. [User Guide](https://docs.maiagent.ai/build/ai-gateway)
* **\[Shared] Gemma 4 supports thinking depth**: Gemma 4 can now have thinking turned on or off in assistant settings, with support for Bedrock Mantle, OpenRouter, and self-hosted vLLM. [User Guide](https://docs.maiagent.ai/build/ai-gateway)
* **\[Shared] Model access page redesign**: The model access page has been redesigned with capability scores, a comparison chart, and bulk toggles; "Disable Chinese models" now keeps those models visible but unselectable, and the organization overview shows when it is on. [User Guide](https://docs.maiagent.ai/build/ai-gateway)
* **\[Shared] Role assignment dialogs support removal and select-all across pages**: Role assignment dialogs can now directly remove existing AI assistants, tools, skills, chat platforms, knowledge bases, databases, and members, and both lists support select-all beyond the current page. [User Guide](https://docs.maiagent.ai/org/roles)
* **\[Shared] Organization owners can edit member names**: Organization owners can change the name a member displays within the organization. [User Guide](https://docs.maiagent.ai/org/overview)
* **\[Shared] Hide built-in official assistants**: Dropdowns in transfer lists, Agent Teams, form assistants, contact filters, AI assistant monitoring, and quota allocation now offer an option to hide built-in official assistants. [User Guide](https://docs.maiagent.ai/build/agent-builder)
* **\[Shared] Parser menu shows pricing and use cases**: The parser menu adds a Vision economy tier, per-page prices, and use cases; usage records show page counts, and the scanned-page OCR billing item description is now generic. [User Guide](https://docs.maiagent.ai/build/ai-gateway)

#### 🚀 Core Performance Optimization

* **\[Shared] Offline parser upgraded with faster scanned-document recognition**: After the MaiAgent offline parser upgrade, scanned-page recognition recall rose from 61.7% to 94.0%, and processing time per page dropped from 32.7 seconds to 7.8 seconds.
* **\[Shared] Azure and OpenAI prompt cache hits after memory recall adjustment**: In multi-turn conversations on Azure OpenAI and OpenAI with long-term memory enabled, history can now reliably hit the prompt cache.
* **\[Shared] Faster loading of the usage records list**: The credit usage records list for large organizations is aggregated only once and uses an index, so the page no longer takes a long time to open.
* **\[Shared] AI assistant tool declarations loaded on demand**: Tool declarations for files and history lookup are no longer attached when the conversation has no files or does not need history, saving about 2.5K tokens on 73% of Agent replies.
* **\[Shared] Faster first sandbox call each turn**: The sandbox isolation probe wait was cut from 3 seconds to 1.5 seconds, so the first sandbox call each turn waits 1.5 seconds less.
* **\[Shared] Faster loading of contact usage statistics**: The contact tab of usage statistics now groups by ID and looks up names only for the current page, so organizations with many contacts no longer wait 3 to 22 seconds.
* **\[Shared] Faster organization used-space query**: The used upload space query on the organization info page now uses a dedicated index and loads faster.
* **\[Shared] Pre-reply quota estimation no longer blocks replies**: When the quota estimate cache expires, it is now recomputed in the background so user replies no longer wait.

#### 😊 User Experience and Interface Improvements

* **\[Shared] Connector request notification emails now sent in Traditional Chinese**: Connector request notification emails are now sent in Traditional Chinese.
* **\[Shared] Updated shared space permission descriptions**: The descriptions of the shared space create and manage permissions now read "Create and delete shared spaces, and edit space names and authorized roles".
* **\[Shared] Invitation API expiry supports two formats**: The invitation expiry and quota expiry fields of the invitation API now both accept ISO 8601 and millisecond timestamps.
* **\[Shared] Model access page still lists Chinese models marked as blocked when they are disabled**: With "Disable Chinese models" on, the model access page still lists models from Chinese platforms and marks them as blocked.
* **\[Shared] Organization usage cards in four columns**: When the organization overview shows the file library space card, usage cards are now laid out in four columns on wide screens.
* **\[Shared] Billing description adds AI Gateway voice scenario**: The voice billing item description now covers usage of the AI Gateway voice API.
* **\[Shared] Role permission governance entries shown for all organizations**: All organizations can now see the role permission governance entries.
* **\[Shared] Support center entry now uses the new domain**: The support center entry now points to the new MaiAgent Helpdesk domain, and the hint text no longer includes the product name.

#### 🔐 Security and Governance Enhancements

* **\[Shared] OAuth password grant now covered by LDAP authentication**: In organizations with LDAP enabled, exchanging a token through the OAuth password grant now also goes through LDAP authentication.
* **\[Shared] MCP tool attachment ownership check**: MCP tool calls now check whether an attachment belongs to the current conversation, and attachments from other conversations are blocked.
* **\[Shared] Fix for member removal happening at the same time as SSO login**: A removed member's removal is no longer reversed even if they sign in through SSO at the same time.
* **\[Shared] Telegram Bot Token and Text2SQL connection strings stored encrypted**: Telegram Bot Tokens and Text2SQL external database connection URLs are now stored encrypted, so plaintext no longer appears in backups.
* **\[Shared] Default roles no longer grant file library permissions**: Default roles no longer grant file library permissions; to open access to members, create a custom role.
* **\[Shared] Markdown links no longer execute javascript**: Markdown links in chat and Canvas are no longer executable javascript links, and the preview console accepts messages only from the preview window.

#### 🛠️ Bug Fixes

* **\[Shared] OpenAI direct GPT-5.6 Sol stays available**: The OpenAI direct GPT-5.6 Sol model is no longer disabled by a scheduled retirement after deployments, and AI assistants using it are no longer automatically moved back to the Azure endpoint.
* **\[Shared] OpenAI models on Bedrock support token-by-token streaming**: For AI assistants using GPT-6.1 Sol and GPT-5.6 Sol on Bedrock, answers now appear token by token instead of all at once after the whole response completes.
* **\[Shared] Fallback models are billed by the model that actually answered**: Conversation replies now deduct credits based on the model and unit price that actually answered, instead of carrying over the primary model when a fallback group takes over.
* **\[Shared] Answers blocked by the model provider no longer deduct credits**: When using platform models, a call whose answer is blocked by the provider's content moderation no longer deducts credits.
* **\[Shared] API documentation aligned with actual behavior, and API Key now recorded**: Corrected five API documentation descriptions that did not match actual behavior, and API call logs now show which API Key was used.
* **\[Shared] Ask AI long tasks no longer fail because of expired credentials**: Credentials are now refreshed automatically during Ask AI long tasks, so 401 errors no longer appear after 10 minutes, and retries stop after consecutive failures.
* **\[Shared] API key requests with a non-object body now return 400**: When sending a message with an API key, a body that is not a JSON object now returns 400 instead of 500.
* **\[Shared] Files that fail Vision parsing no longer deduct credits**: If a knowledge base file ultimately fails after Vision parsing, the customer is no longer charged credits for that run.
* **\[Shared] Claude prompt cache markers are now sent**: For Claude models on Vertex or direct Anthropic, prompt caching now actually takes effect once enabled.
* **\[Shared] Text-to-SQL deletion and re-upload fixes**: Deleting a Text-to-SQL data file no longer wrongly clears its own tables, and interrupted files can be re-uploaded and completed.
* **\[Shared] Vision Parser resume and billing fixes**: Resuming an interrupted Vision Parser run now picks up progress correctly and records usage completely.
* **\[Shared] One-off calls no longer send cache markers**: One-off calls such as memory consolidation and question rewriting are no longer billed at the cache-write price.
* **\[Shared] Real-time information labeled as platform information**: After tool calls, AI assistants no longer mistake real-time information appended by the platform for something the user said.
* **\[Shared] Credit records fixed for non-message-billed turns**: For Team, restart-interrupted, and no-charge turns, tool charges are now recorded on that turn's assistant reply.
* **\[Shared] Web search and map tools without a key no longer deduct credits**: When no key is configured, web search and map tools are recorded as failed and no longer deduct credits.
* **\[Shared] Multi-item citation matching for the web fetch tool**: The global web fetch tool no longer wrongly reports "not found" for multi-item questions, and fetch failures are recorded as failed calls.
* **\[Shared] Vision parsing rate-limit handling**: When many files use Vision Parser at once, concurrency is controlled by quota, and rate-limited pages are queued automatically and resumed.
* **\[Shared] Long-term memory is no longer answered as a new user message**: When an assistant with long-term memory answers after a tool call, the memory attached by the platform is no longer treated as a new question from the user.
* **\[Shared] OpenAI and Azure cache-write credits follow provider prices**: Cache-write costs for OpenAI and Azure models are now counted in credits according to the provider's published prices.
* **\[Shared] SSL relaxed-domain list applied to image downloads**: The SSL relaxed-domain list now also applies to image downloads for text-to-image reference images and image-to-video, so storage using its own certificates can be accessed normally.
* **\[Shared] Back office auto-creates cache-write pricing for OpenAI / Azure models**: When an OpenAI or Azure model is added in the back office, a credit pricing rule for cache writes is created with it.
* **\[Shared] On-premises upgrades skip official Agents when a global tool is missing**: During an on-premises upgrade, if the site lacks a tool an official Agent needs, that Agent is skipped and the missing items are listed, instead of the upgrade failing and rolling back.
* **\[Shared] Unauthorized contacts can get MCP tool authorization links**: With tool search enabled, unauthorized contacts can receive authorization links for MCP tools, and the reminder names the services not yet authorized.
* **\[Shared] Clear message when registering with an already-registered email**: Registering or accepting an invitation with an email that is already in use now shows "An account has already been registered with this email" instead of an error.
* **\[Shared] Tools work when headers carry Chinese contact names**: API and MCP tools now work normally when headers carry Chinese contact names.
* **\[Shared] Fixed Responses API model reply failures**: For assistants using Responses API models such as GPT-6 Luna, the tool loop no longer fails the entire reply when it encounters an empty text segment.
* **\[Shared] API documentation fills in field descriptions for invitations and new-member quotas**: The API documentation now includes request and response formats for unified invitations, default quotas for new members, and quota settings.
* **\[Shared] OpenAI direct GPT-6 Luna / Sol can adjust thinking depth**: OpenAI direct GPT-6 Luna and Sol can now adjust thinking depth in the assistant settings page and MaiGPT.
* **\[Shared] API key allowlists that contain only own models are no longer treated as unrestricted**: When an API key's AI Gateway allowlist lists only the organization's own models, it is no longer mistaken for unrestricted.
* **\[Shared] Models past their retirement date no longer appear in the menu**: Platform models past their retirement date no longer appear in the model menu and cannot be newly bound; a retirement status column has been added to the back office.
* **\[Shared] AWS Marketplace usage reporting now has a schedule**: AWS Marketplace usage is now reported automatically every hour, with retries on failure and a daily reconciliation.
* **\[Shared] Fixed the unnamed "Unknown" in usage statistics**: The "Unknown" value in the "Chat platform / Member" columns of usage statistics now has a specific name and an explanation.
* **\[Shared] Fixed the blank screen when a page fails to load after an update**: When a page module fails to load, the page now reloads automatically once instead of getting stuck on a blank screen.
* **\[Shared] Fixed icons failing to load in offline environments**: Icons are now bundled into the frontend, so offline sites no longer fetch icons from external services.
* **\[Shared] Fixed the answer being cleared when confirming Zhuyin candidates in Safari**: In Safari, confirming Zhuyin candidate selection with Enter no longer clears the custom answer in the question card.
* **\[Shared] Fixed the account menu not expanding on phones and tablets**: Tapping the avatar on phones and tablets now expands the account menu so users can log out.

</details>

### 🗂️ Full Update (by Product Line)

<details>

<summary>Work (43 items)</summary>

#### ✨ New Features

* **\[Work] Dedicated panel for MaiGPT document translation**: MaiGPT now presents document translations in a dedicated translation card, translating paragraph by paragraph and showing the source file details along with the source and target languages. [User Guide](https://docs.maiagent.ai/work/maigpt)
* **\[Work] Schedules auto-disable after consecutive failures, and short intervals require a stop condition**: A schedule is disabled automatically, and its creator notified, when consecutive failures reach the limit or the end date arrives. Schedules that run more often than once an hour must have a stop condition. [User Guide](https://docs.maiagent.ai/work/overview)
* **\[Work] MaiGPT organization sharing**: A conversation shared with the organization follows the original conversation, and its creator can change who it is shared with and keep chatting at any time. [User Guide](https://docs.maiagent.ai/work/maigpt)
* **\[Work] Desktop Cowork local browser tool integration**: MaiGPT's Desktop Cowork can use local browser tools, and the cloud browser is no longer offered alongside it when a local browser is available. [User Guide](https://docs.maiagent.ai/work/maigpt)
* **\[Work] Export translations as Word files that keep the original formatting**: After a Word document is translated, you can download the translated version as a Word file that keeps the original formatting. [User Guide](https://docs.maiagent.ai/work/maigpt)
* **\[Work] Layout-preserving PDF export can reuse Canvas translations**: When exporting a PDF that keeps the original layout, you can choose to reuse the content already translated and edited in Canvas. [User Guide](https://docs.maiagent.ai/work/maigpt)
* **\[Work] Quality validation before document translation export**: Before a translated document is exported, its text and layout quality are checked, and files that fail are not offered for download as successful results. [User Guide](https://docs.maiagent.ai/work/maigpt)
* **\[Work] In-place translation of text in images embedded in documents**: Text inside images embedded in PDF and Word documents can now be translated in place. [User Guide](https://docs.maiagent.ai/work/maigpt)
* **\[Work] Resumable uploads in the file library**: The file library adds a resumable upload mechanism, so an interrupted upload can continue from where it stopped, with integrity verification. [User Guide](https://docs.maiagent.ai/work/file-library)
* **\[Work] Meeting record live speech engine details and live translation model choice**: The live speech engine for meeting records shows its country of origin, data location, and supported spoken languages, and live translation lets you pick a model from three tiers. [User Guide](https://docs.maiagent.ai/work/meeting-records)

#### 🚀 Core Performance Improvements

* **\[Work] Prompt cache continues after a Canvas handoff**: After a handoff, the Canvas assistant reuses the main assistant's tools and system instructions so the prompt cache carries over, cutting credit consumption by about 4.3%.
* **\[Work] Shorter wait for responses when AI Memory is labeled**: The AI Memory location estimate now runs outside the main model call, and the limit on how much memory is included uses a new counting method.
* **\[Work] Faster conversation visibility queries**: When members view the conversation list, visibility is now determined with a more efficient query, so large organizations load faster.
* **\[Work] Unified tool loading mode for MaiGPT**: The tool loading mode of every MaiGPT is now decided by the platform template, and legacy MaiGPTs no longer load all tools on every turn.

#### 😊 User Experience and Interface Improvements

* **\[Work] File library browse page shows version history and usage**: The file library browse page adds library-wide version history, used storage, text previews, and file creator information.
* **\[Work] Handling and notification for schedules with insufficient credits**: A schedule that hits insufficient credits no longer counts toward its run total, and its creator is notified when it is disabled for insufficient credits or for reaching its run limit.
* **\[Work] MaiGPT workspace removes the 20-file limit per conversation**: The MaiGPT workspace can now read files in a single conversation without the 20-file cap.
* **\[Work] Slides skill accepts typed answers for attachment and template questions**: When the AI slides skill asks what an attachment is for and which template to use, users can answer by typing text directly.
* **\[Work] More accurate live translation of speech-recognition fragments**: Meeting record live translation now treats its input as speech-recognition fragments, cutting the average number of mistranslated segments from about 19 to about 15.

#### 🔐 Security and Governance Enhancements

* **\[Work] Default roles no longer carry file library permissions**: File library access is now granted through explicitly assigned custom roles, and file library permissions on default roles no longer take effect.

#### 🛠️ Bug Fixes

* **\[Work] AI assistant file changes run only after user confirmation**: In organizations that have not enabled skipping confirmation, MaiGPT first lists its planned file changes and only creates folders, renames, and moves files after the user confirms.
* **\[Work] Repeated content in audio transcripts is collapsed automatically**: After an uploaded audio file is recognized, the transcript no longer contains hundreds of repetitions of the same words, and a repeated phrase is kept at most 3 times.
* **\[Work] Audio transcription moves to Groq because of the OpenAI model retirement**: WhisperOpenAI audio transcription for knowledge bases and meeting records is now handled by a Groq model, and segment timestamps are kept.
* **\[Work] Stability fix for user memory extraction**: When MaiGPT memory organizes facts about a user, it no longer mistakenly answers instructions found in messages, and extraction results are saved normally.
* **\[Work] Replies save correctly when searching conversation messages with a date range**: After asking the assistant to search conversation messages by date range, the answer is saved to the conversation history normally and does not disappear on refresh.
* **\[Work] MaiGPT map canvas no longer appears as raw text**: When you ask MaiGPT to mark locations on a map, the map is shown in the side panel instead of a block of raw data appearing in the conversation.
* **\[Work] Agents in the file workspace can create folders and complete rename and move together**: MaiGPT and file workspace Agents can create new folders, and renaming and moving the same file now happen together.
* **\[Work] Slide export images support the SSL relaxed domain list**: Image retrieval during slide export now applies the SSL relaxed domain list, so on-premises sites with TWCA certificates can export correctly.
* **\[Work] Missing download-link files now suggest available files**: When the AI assistant requests a download link for a file that does not exist, it sees the list of downloadable files in that conversation and can choose another one.
* **\[Work] AI Slides are built to the page count the user types**: When AI Slides asks for the number of pages, a number the user types in is used directly as the page count.
* **\[Work] PDF translation supports PDFs printed from a browser**: PDFs printed from a browser can now produce a layout-preserving bilingual PDF, and a clear reason is shown when a file cannot be processed.
* **\[Work] PDF translation reads multi-column layouts in order**: Multi-column PDFs are translated column by column, and sentences split across line breaks are merged into complete paragraphs.
* **\[Work] Canvas agent default prompt adds Mermaid line-break rules**: The line-break rules for node text in Mermaid flowchart canvases are now part of the default prompt; clear the custom value in the admin console to apply the latest rules.
* **\[Work] AI Slides verifies images as each page is written**: When AI Slides writes a page containing generated images, it checks on the spot whether the images are displayed, and reports clearly if they are not instead of reporting completion.
* **\[Work] Image-to-video can use images uploaded in the conversation**: Image-to-video can use images uploaded in the conversation directly, and they are no longer blocked as external URLs.
* **\[Work] Tools with a runtime parameter no longer cause reply failures**: Tools with an execution-environment parameter, such as Desktop Run Script, now run and finish normally instead of showing "The system is temporarily unable to process this request."
* **\[Work] MaiGPT desktop browser tools stay enabled after deployment**: The Desktop local browser tools are enabled and bound to MaiGPT automatically after deployment, with no need to turn them on manually.
* **\[Work] Fixed Desktop tools being rejected because of optional parameters**: Desktop tools with optional parameters are no longer rejected, and parameters the model left blank are not forwarded.
* **\[Work] Fixed being unable to rename a meeting right after a long recording ends**: Meeting names can be changed immediately after a long recording ends, even while it is still being converted, and edits no longer overwrite the conversion result.
* **\[Work] PDF translation restores the Layout tab**: The Layout tab in the PDF translation canvas is shown again, so paragraph positions can be viewed.
* **\[Work] File library checks content before applying pending suggestions**: Applying a pending file library suggestion now confirms the content has not changed, so content the user did not preview is never applied.
* **\[Work] Fixed file size of exported translated PDFs**: When document translation exports a PDF in the original layout, the number of image labels no longer multiplies the file size.
* **\[Work] Images generated before AI Slides are created can be used in Insert Image**: Images that MaiGPT generated before creating a slide deck now also appear in the Insert Image drawer and the deck's folder.

</details>

## v2026.10.17 <a href="#v2026-10-17" id="v2026-10-17"></a>

***

### 📌 Highlights

* **\[Shared] Better cache reuse in long AI assistant conversations**: AI assistants using Claude models, such as MaiGPT, previously could rewrite the entire cache when a new question was asked in a long, multi-turn conversation. The system now adds an extra cache anchor at the previous question, so subsequent questions only require processing newly added content. Conversation history reuses the existing cache, reducing repeated work.

### 🗂️ Full Update (by Product Line)

<details>

<summary>Shared (1 item)</summary>

#### 🚀 Core Performance Improvements

* **\[Shared] Better cache reuse in long AI assistant conversations**: AI assistants using Claude models, such as MaiGPT, previously could rewrite the entire cache when a new question was asked in a long, multi-turn conversation. The system now adds an extra cache anchor at the previous question, so subsequent questions only require processing newly added content. Conversation history reuses the existing cache, reducing repeated work.

</details>

## v2026.10.16 <a href="#v2026-10-16" id="v2026-10-16"></a>

***

### 📌 Highlights

* **\[Serve] Fixed pagination count timeouts on the conversation history page**: Members who are not organization owners could wait tens of seconds or encounter a timeout when opening the admin conversation history page in organizations with many inboxes. The count now reads only conversations belonging to their organization, reducing the worst-case duration from about 35 seconds to less than 0.1 seconds.

### 🗂️ Full Update (by Product Line)

<details>

<summary>Serve (1 item)</summary>

#### 🛠️ Bug Fixes

* **\[Serve] Fixed pagination count timeouts on the conversation history page**: Members who are not organization owners could wait tens of seconds or encounter a timeout when opening the admin conversation history page in organizations with many inboxes. The count now reads only conversations belonging to their organization, reducing the worst-case duration from about 35 seconds to less than 0.1 seconds.

</details>

## v2026.10.15 <a href="#v2026-10-15" id="v2026-10-15"></a>

***

### 📌 Release Highlights

* **\[Work] Personal AI memory switch**: The AI Memory page in your profile now includes a “Let the assistant remember me” switch, configured separately for each organization. Turning it off stops new memories from being saved and existing memories from being used in conversations. Saved memories remain available to view and manage. [User Guide](https://docs.maiagent.ai/work/maigpt/ai-memory)
* **\[Work] Workspace file deletion approval**: The file workspace toolbar now shows the number of pending deletion requests. Owners can use version history to review the filename, requester, source conversation, and deadline, then approve or reject each request submitted by an Agent on a member's behalf. [User Guide](https://docs.maiagent.ai/work/file-library)

### 📌 Highlights

* **\[Build] Back-office chats respect assistant tag filters**: When reference documents are left unchanged, back-office chats use the assistant's existing tag filters, matching the default retrieval scope of Web Chat; once documents are explicitly selected, retrieval still follows that selection.
* **\[Shared] Inline citations no longer delay the start of an answer**: With inline citations enabled in standard answer mode, answer generation no longer waits for citation data to finish sending, reducing the wait before the answer starts; citations can be added after the text begins. When integrating through the API, allow citation information to arrive after part of the text.
* **\[Serve] Batch deletion of Web Chat conversation history**: Web Chat conversation history now has an edit mode where you can select several or all currently loaded conversations and delete them in one batch; if some deletions fail, the numbers of successes and failures are reported and the failed items are kept for retry. [User Guide](https://docs.maiagent.ai/serve/daily-operations/reply)

### 🗂️ Full Update (by Product Line)

<details>

<summary>Serve (12 items)</summary>

#### ✨ New Features

* **\[Serve] Batch deletion of Web Chat conversation history**: Web Chat conversation history now has an edit mode where you can select several or all currently loaded conversations and delete them in one batch; if some deletions fail, the numbers of successes and failures are reported and the failed items are kept for retry. [User Guide](https://docs.maiagent.ai/serve/daily-operations/reply)

#### 😊 User Experience and Interface Improvements

* **\[Serve] Service impact warning before switching Copilot mode**: Selecting the Copilot "Every incoming message" mode now clearly warns that the AI will not reply automatically, hand over to a human automatically, or speak in web calls, reminding you to confirm that support staff can reply promptly.
* **\[Serve] Consistent Traditional Chinese wording and multilingual prompts in Web Chat**: Web Chat's Traditional Chinese now uses wording common in Taiwan, with "branch conversation" used consistently; prompts for the home screen, installation, and connection errors now follow the current language.
* **\[Serve] Notice before attachments are sent as links on external channels**: When a file type that LINE or Instagram cannot send directly is attached to a conversation, a notice before sending explains that a download link will be sent instead, so support staff know what the customer will receive.
* **\[Serve] Hand back to AI from any tab**: "Hand back to AI Agent" for support conversations can now be used from any tab, so supervisors or colleagues no longer need to assign the conversation to themselves and switch to "Mine" first; the status bar also wraps on narrow screens so the button stays visible.

#### 🔐 Security and Governance Enhancements

* **\[Serve] Interactive cards restrict external connections by default**: When no allowed origins are configured for Web Chat's MCP interactive cards, external websites are no longer allowed by default; services that need external access must have allowed origins explicitly configured by an administrator.
* **\[Serve] Permission check before adding a channel**: Without the "Allow adding conversation platforms" permission, the connect-channel entry is disabled with an explanation, and opening the add page directly is also blocked up front, so you do not fill in the form only to find it cannot be saved.

#### 🛠️ Bug Fixes

* **\[Serve] Authorization cards no longer reuse another contact's state**: After Web Chat switches contacts or clears the identity, the previous contact's connection state is cleared first; if loading fails, the authorization entry for the current identity is shown instead of wrongly showing someone else's "Connected".
* **\[Serve] Complete reply endings**: After Web Chat receives a reply that has finished saving, it fills in any text missing from the end of the stream and verifies the order, so the live answer matches what you see after a refresh.
* **\[Serve] Support conversation switching no longer shows a blank view**: 1. Switching tabs while conversation messages are still loading now updates correctly instead of staying blank; 2. List previews strip Markdown symbols, and the loading-complete text correctly refers to the conversation or message; 3. The Copilot manual mode description now matches the actual entry point.
* **\[Serve] Human-handoff waiting state syncs after reconnecting**: When Web Chat reconnects and catches up on events, it also handles this round's cancelled handoffs, agent takeovers, or hand-backs to AI, clearing stale waiting states and unlocking the input box so there is no need to wait for the queue to time out.
* **\[Serve] Reloading settings keeps queue-period drafts**: When the human-handoff settings are reloaded, unsaved queue-period edits are kept, including content typed while a save is pending; fields you have not changed are still updated to the latest server values.

</details>

<details>

<summary>Build (18 items)</summary>

#### 😊 User Experience and Interface Improvements

* **\[Build] Schedule action and run status descriptions**: The schedule toggle tooltip now states exactly which action is about to run, and run details and filters consistently show "Partially succeeded", making runs that did not fully complete easy to spot.
* **\[Build] Hook logs show what was checked**: Hook run logs, filters, and details now show the trigger point, action, and match result in readable text, and a new checked-content field distinguishes sources such as user messages and knowledge base retrieval content.
* **\[Build] Satisfaction charts include feedback counts**: The satisfaction chart tooltip now shows the ratio together with the numbers of likes and dislikes, available in hourly, daily, and monthly statistics, making it easier to tell no feedback apart from actual dissatisfaction.
* **\[Build] Disabled databases show an enable prompt**: The database selector for an AI assistant now lists disabled databases marked as unselectable and prompts you to enable them on the Databases page first; databases that are linked but disabled, and the settings page, also explain that the assistant cannot query them for now.
* **\[Build] New assistants truncate overly long tool results by default**: When a new AI assistant is created in the back office, "Tool result truncation" is on by default, matching the platform default; it can still be adjusted in advanced settings as needed.
* **\[Build] Agent Marketplace remembers the last tab**: The Agent Marketplace remembers the tab you last viewed in the same browser and returns to it when you re-enter or refresh; after clearing cookies or switching devices or browsers, the default tab is used.
* **\[Build] Skills show as unavailable outside Agent mode**: When the assistant's answer mode is not Agent mode, the skill settings show "Available only in Agent mode" and provide a link to the answer mode settings, so skills that are not in effect are not mistaken for a channel problem.
* **\[Build] FAQ import shows added and skipped counts**: The FAQ batch upload button is now "Upload and add", clearly stating that rows with the same question are skipped and their answers are not overwritten; the completion notification and import log show the added, skipped-duplicate, and error counts separately.
* **\[Build] Agent Marketplace search and management in one place**: The Agent Marketplace now gathers the current tab's information, management entry points, and search in the title bar, keeps category and favorites filters above, and scrolls only the Agent card area.

#### 🛠️ Bug Fixes

* **\[Build] Back-office chats respect assistant tag filters**: When reference documents are left unchanged, back-office chats use the assistant's existing tag filters, matching the default retrieval scope of Web Chat; once documents are explicitly selected, retrieval still follows that selection.
* **\[Build] Confirm available roles when adding skills and tools**: When adding a skill, tool, or Flex Message interface tool without having looked at its permission settings, saving first takes you to the permissions tab with a reminder to confirm, so a newly created resource is not visible only to its owner.
* **\[Build] Tool save failures show the specific reason**: When saving a tool or interactive interface tool is rejected, the actual reason, such as a duplicate name, is now shown and duplicate error pop-ups are avoided, so you can fix the field directly.
* **\[Build] Agent previews still play with reduced motion**: Agent marketplace previews now still play the demo conversation and walkthrough message by message when "Reduce motion" is enabled, and can be replayed, while skipping typing animations and smooth scrolling.
* **\[Build] Crawl failure status labeled correctly**: When a web crawl fails, the task list, details, page list, and schedule history now all show "Failed", consistent with the filters, and the specific failure reason can still be viewed.
* **\[Build] Evaluation reports show the correct assistant**: The "AI Assistant" field in the single-question evaluation detail report now shows the assistant actually evaluated in that run, matching the evaluation settings; when data has not loaded yet, an empty value is shown to avoid confusion.
* **\[Build] AI search results and summary use the same query**: Knowledge base AI search now generates results, summaries, and search history from the text as submitted; submitting a different keyword mid-search starts a new query, and older results that arrive later no longer overwrite the new search.
* **\[Build] Assistant settings no longer flag your own updates**: After adding or removing knowledge bases, tools, connectors, skills, or databases in assistant settings, you can now save the name and instructions right away; your own instant saves are no longer mistaken for updates by others, while genuine concurrent edits are still checked.
* **\[Build] Tool editing no longer mixes in earlier responses**: After the tool editing window is closed or you switch tools, late-arriving content, authorization status, and tool lists from the previous round are discarded, so Tool A's settings are never written into Tool B.

</details>

<details>

<summary>Shared (26 items)</summary>

#### 🚀 Core Performance Improvements

* **\[Shared] Inline citations no longer delay the start of answers**: When inline citations are enabled in standard answer mode, answer generation no longer waits for citation data to finish being sent, reducing the wait before an answer begins; citations can be added after the text starts. When integrating through the API, allow citation information to arrive after part of the text.

#### 😊 User Experience and Interface Improvements

* **\[Shared] Announcement reminders and read behavior are consistent**: 1. Feature announcements are only shown on login, load, or page change, and closing one no longer pops up the next in succession; 2. Organization announcements can be presented according to the chosen reminder method, and brief notices start their countdown only after the dialog is closed; 3. Clicking an announcement link marks it as read, so it is not shown again after a refresh.
* **\[Shared] Admin console guidance matches page names**: 1. Login source descriptions and the new-member invitation template now reflect the actual steps; 2. Call Center guides you based on the pages available instead of always landing on number settings; 3. Tab titles for Credit export and chat platforms match their actual content.
* **\[Shared] Real-time voice call billing descriptions localized**: "Real-time voice call duration" in billing methods now has names and descriptions in each interface language, making it easier to see how call time counts toward usage.
* **\[Shared] Email invitation validity description corrected**: Email invitations now state that they "remain valid until the recipient joins or the invitation is revoked" and no longer display an invitation expiry; invitation links can still have an expiry date set, and the selected period is kept when you switch invitation methods.
* **\[Shared] Organization settings descriptions match what is editable**: When assigning an AI assistant, the interface now explains that its chat platform will be assigned along with it; announcement and model invocation permissions are read-only for non-owners with an explanation, reducing permission errors that only appear after taking an action.
* **\[Shared] API key error response hints corrected**: The API key management drawer now shows the correct hints: subsequent calls with a disabled key return 401, and models that are not enabled return 404, making it easier to troubleshoot integration issues based on the actual responses.
* **\[Shared] Default resource permission assignments are clearly visible**: When creating a resource, the permission settings now mark the default option that has been applied and list a summary of role assignments after creation; non-owners also see an explanation of the assignable scope, so it is not mistaken for authorizing the whole organization.
* **\[Shared] SSO activation confirmation notes existing passwords still work**: The confirmation dialog for enabling SSO now states clearly that members who have set a password can still sign in with their account and password, helping administrators plan account management around the actual sign-in rules.
* **\[Shared] Setting options and integration descriptions aligned**: 1. Message Webhooks now offer only POST, PUT, and PATCH, and explain the push direction for visitor messages and assistant replies; 2. Push logs show the conversation ID correctly, and re-uploading a skill package is offered only for skills created by upload; 3. The description of the knowledge base similarity threshold now matches how filtering actually works.
* **\[Shared] Integration examples match the actual interfaces**: The Source ID example in the admin console now waits for the SDK to load before initializing identity, the API description uses the correct Api-Key header and /api/v1/ path, and the file upload example now includes the required steps.

#### 🔐 Security and Governance Enhancements

* **\[Shared] Canvas preview isolation protection**: The admin console and Web Chat now strengthen isolation for generated canvases, restricting canvas content from accessing sign-in information or manipulating the outer page; interactive preview and PDF export remain available, and unsafe content is sanitized before display.
* **\[Shared] Cloud analytics tracking removed from on-premises builds**: New on-premises builds of the admin console and Web Chat remove the cloud analytics tracking configuration, avoiding the use of cloud analytics services that send events to the platform; this applies after deploying the new on-premises images.
* **\[Shared] Navigation restrictions for unreleased pages**: The production site no longer exposes internal demo pages; when the billing management entry is turned off, opening its URL directly redirects to an available organization page, preventing use of features that are not yet released.
* **\[Shared] Real-time connections no longer carry credentials in the URL**: Real-time connections in the admin console and Web Chat now pass the sign-in credential through authentication data, so it no longer appears in connection URLs and access logs; existing connections are cleaned up on sign-out, and reconnection uses the current signed-in identity.

#### 🛠️ Bug fixes

* **\[Shared] Consistent canvas source file download formats**: Mermaid diagrams downloaded from any entry point now consistently use .mmd; map canvases offer only the coordinate source file in .json, and downloading the map source file in the admin console, which previously did nothing, is fixed.
* **\[Shared] On-premises onboarding shows complete contact information**: When an on-premises site with MaiAgent branding has no support email configured, a default contact email is now provided; for custom brands that do not provide an email, the empty contact entry is hidden so no email opens without a recipient.
* **\[Shared] Canvas export checks the current content**: Canvas export now verifies the content and preview version at the moment you click download; if you switch canvases or recompile while waiting, that export is aborted to avoid downloading the wrong version. When exporting a PDF from code mode in Web Chat, the preview for that version is completed first.
* **\[Shared] Currency symbols are no longer typeset as math**: NT$, US$, HK$ and ordinary amounts in replies, citations and document previews are shown as written, and no longer turn italic or show a math-loading indicator when several amounts appear in one paragraph.
* **\[Shared] On-premises documentation links open the correct help pages**: In the on-premises admin console, the model selection tips and role instruction help link to the matching documentation pages; when no documentation site is configured, a default help entry is used so that invalid URLs are never opened.
* **\[Shared] Multiple cards in one reply appear in real time**: When one reply produces several interactive cards or several sets of quick replies, the admin console and Web Chat now keep each one in order instead of overwriting them, so the live display matches what you see after a refresh.
* **\[Shared] Model tool-calling capability is labeled correctly**: The model menu now shows the "Function calling" icon and description based on each model's actual capabilities in the model catalog, so models that support tool calling are correctly recognized.
* **\[Shared] Identity sync permissions and quota alerts aligned**: The identity sync page is now controlled by its own dedicated permission, and the related permission names are completed; the "Low balance" threshold on personal and role quota cards follows the organization's alert settings, so cards and notifications agree.
* **\[Shared] File-writing progress and error messages display correctly**: When an assistant calls other tools while writing a file, the Web Chat file card no longer stays blank and stuck, and the admin console no longer exposes the file contents; errors during generation are shown as ordinary conversation text instead of being hidden inside the file content.
* **\[Shared] Links to newly joined organizations open directly**: When you open an organization notification or Email link and the target is not in the cached list, the system first reloads your organization memberships before deciding, so a newly joined organization can be entered directly without refreshing the whole page first.
* **\[Shared] Credit details use the latest filter result**: When you switch Credit usage filters quickly, the list uses only the latest query result so transactions from other categories are not mixed in; the loading state also reflects only that list's own request.

</details>

### 🗂️ Full Updates by Product Line

<details>

<summary>Work (24 items)</summary>

#### ✨ New Features

* **\[Work] Personal AI memory switch**: The AI Memory page in your profile now includes a “Let the assistant remember me” switch, configured separately for each organization. Turning it off stops new memories from being saved and existing memories from being used in conversations. Saved memories remain available to view and manage. [User Guide](https://docs.maiagent.ai/work/maigpt/ai-memory)
* **\[Work] Workspace file deletion approval**: The file workspace toolbar now shows the number of pending deletion requests. Owners can use version history to review the filename, requester, source conversation, and deadline, then approve or reject each request submitted by an Agent on a member's behalf. [User Guide](https://docs.maiagent.ai/work/file-library)

#### 😊 User Experience and Interface Improvements

* **\[Work] Live translation models display their providers**: The meeting live translation model menu shows the provider beside each model name. Both remain visible when the menu is collapsed, making the selected model easier to identify.
* **\[Work] Localized file space names and paths**: File library search results and the MaiGPT workspace picker display personal space, organization space, and temporary storage names in the interface language. Custom space and folder names remain unchanged, and the picker guidance now reflects current file usage.
* **\[Work] Generated files appear immediately with links to the library**: AI-generated files appear in the output panel right away, with a processing status while they are saved to the file library. Once archived, the original item provides access to the file's location.
* **\[Work] Reorganized connector and skill menus**: The composer's “+” menu lists only authorized, available connectors. Authorization and adding connectors are handled through “Manage connectors,” and empty skill or tool groups are hidden. This also applies to the classic MaiGPT layout.
* **\[Work] Clearer file deletion and restore messages**: 1. Deletion confirmations identify the file and consequences, and restore conflicts show “Restore failed” with a reason; 2. Deeply nested folder names are truncated to leave room for actions; 3. Shared spaces now include relevant guidance and empty states to clarify which actions are available.
* **\[Work] Organization sharing list pagination and layout updates**: The organization sharing page separates the data area from the pagination bar at the bottom, with 10, 20, 50, or 100 items per page. The title and actions stay in place while you browse longer sharing histories.
* **\[Work] File library actions stay in place**: The file library pins the title, search, upload, path, and toolbar above the data area. The file list scrolls independently so you can keep managing files from the same position.
* **\[Work] Meeting list pagination and horizontal browsing updates**: The meeting records list keeps its header and filters pinned. It defaults to 20 items per page, with options for 10, 20, 50, or 100. A floating horizontal scrollbar and a compact action menu make wide tables easier to browse.
* **\[Work] Consistent list menus and confirmations**: List action menus adjust their opening position to fit the available space, and deletion uses consistent confirmations and result messages. File renames can be undone from the action notification, with an explicit message when undo is no longer available.
* **\[Work] Presentation shortcut and source selection updates**: The presentation shortcut in classic MaiGPT now uses a fixed length of 10 slides, and the panel closes after confirmation. Sources supplied by templates appear in the composer alongside skills and tools, grouped by count when space is limited.

#### 🛠️ Bug Fixes

* **\[Work] Folders follow list and thumbnail view changes**: In the file library's list view, folders appear at the top of the same table as files, and clicking a folder name opens it. Thumbnail view keeps folder cards, and Select All still selects only files.
* **\[Work] Attachment expiry countdown displays correctly**: The conversation attachment list correctly shows the remaining retention days and marks expired attachments, helping you handle files before their retention period ends.
* **\[Work] Full model names remain visible**: The MaiGPT model menu adjusts its width and narrow-screen wrapping to prioritize complete model names. Hover over a model to see its full name and provider.
* **\[Work] Clear loadable status for attachments**: Attachments that need no parsing consistently show “Ready to load” in file library lists, search results, and details. This covers archived attachments, Agent output, and restored files, making it easier to tell whether AI can use them.
* **\[Work] Mobile file library actions are no longer clipped**: On mobile, the file library places search and action buttons on separate rows and provides a “Spaces and folders” drawer for switching locations. Details, guidance, and version history drawers adapt to the screen width.
* **\[Work] Shared pages load their styles correctly**: When shared pages open from the classic layout, newer sections also load their full styles and organization colors, preventing unstyled lists or dialogs.
* **\[Work] Transcript content survives refresh failures**: If refreshing the same meeting transcript fails, the content already displayed stays visible with a failure message. Switching transcript versions also rechecks the source to avoid showing content from a different version.
* **\[Work] Connector activation notifications open management directly**: Links in connector activation notifications open MaiGPT or Work and show “Manage my connectors.” Notifications for another organization first switch to that organization. Existing notification links also work.
* **\[Work] Knowledge base search preserves selections**: 1. Searching after selecting knowledge bases leaves previously unselected sources unchecked; 2. Links from the knowledge base page to MaiGPT correctly expand the panel; 3. Switching conversation platforms reloads display settings to avoid carrying over knowledge cards from the previous platform.
* **\[Work] Cancelled recordings no longer reopen the microphone**: If a recording is stopped while connecting, ended by the server, or replaced by a new meeting, the old connection cleans itself up when it completes. It no longer opens the microphone, starts backup recording, or stops a subsequently created meeting.
* **\[Work] Switching Agents no longer sends messages to the old conversation**: Switching Agents first clears the previous conversation's send target. If the new conversation fails to load, messages cannot be sent to the earlier conversation, keeping the displayed name consistent with the assistant that actually replies.
* **\[Work] Ask AI history loads with the member's identity**: Admin Ask AI history requests now include the current member's identity so authorized conversations can load, and update authentication information as the login state changes.

</details>

## v2026.10.12 <a href="#v2026-10-12" id="v2026-10-12"></a>

***

### 📌 Release highlights

* **\[Work] Outlook email file (.eml/.msg) parsing**: Knowledge bases and conversation attachments can now process Outlook email files directly: 1. After you upload an .eml or .msg file, the system parses the message body and sender information and lists its attachments; 2. It also extracts content from embedded document attachments, such as quotes, PDFs, and Excel files, so AI can answer questions about their details; 3. Email attachments forwarded through Outlook are recursively expanded; 4. S/MIME-signed or encrypted emails are handled according to their status, and encrypted content that cannot be read is marked; 5. Deleting an email file also removes the attachments extracted from it; 6. TNEF (winmail.dat) files are detected and marked as present without being expanded. Useful for teams organizing customer emails and quote exchanges in a knowledge base or dropping them into a conversation to ask questions. [User guide](https://docs.maiagent.ai/build/knowledge-base)
* **\[Work] Work Agent file workspaces: mount File Library folders on custom and official Agents**: File Library folders can be mounted on Work Agents as workspaces: 1. Custom Agents can have a bound workspace and can allow users to mount their own; 2. Official Agents allow users to mount their own folders; 3. Mounting now supports any nested subfolder, and folder file counts include subfolders. Files modified by an Agent in a workspace can be saved back to File Library. [User guide](https://docs.maiagent.ai/work/file-library)
* **\[Work] Bulk deletion of your conversations**: Select and delete multiple conversations you own from the conversation list in one action, making it faster to clean up a large backlog. [User guide](https://docs.maiagent.ai/work/maigpt)

### 📌 Highlights

* **\[Serve] Conversations automatically return to AI when a human agent does not reply in time**: Conversation platforms gain a new "Auto-return to AI when a human does not reply in time" setting. After a human agent takes over, if no one replies to the visitor's message within the number of seconds you set, the system automatically hands the conversation back to the AI assistant and can send a custom notice at the same time. The timer starts the moment the agent first takes over, a scheduling failure no longer swallows the handoff notification, and messages that arrive at the instant of takeover are not missed. Suited to teams with limited support staff who worry that visitors are left waiting with no one attending. [User Guide](https://docs.maiagent.ai/build/conversation-settings/human-handoff)
* **\[Shared] Bulk-set role permissions and view members' effective permissions**: Organization management gains two permission-governance capabilities: 1. Apply permission settings to multiple MaiAgent roles at once instead of editing them one by one; 2. View the permissions actually in effect for a single member and see which role each permission comes from. The permission screens were also tidied up: role permission names are now fully in Chinese and consistent with the UI, the role permission list is sorted by display name, translations for Hook and conversation platform analytics permissions are completed, and the quota self-check screen now shows the alert threshold. [User Guide](https://docs.maiagent.ai/org/roles)
* **\[Build] AI Gateway adds OpenAI-compatible speech endpoints**: AI Gateway adds OpenAI-compatible speech-to-text and text-to-speech endpoints. Programs that already use the OpenAI speech interface only need to change the URL and key to connect, and usage is deducted from Credits according to actual consumption. [User Guide](https://docs.maiagent.ai/build/ai-gateway)
* **\[Serve] Quoted (replied-to) message content is passed to the AI**: On platforms such as LINE that support quoted replies, when a customer quotes an earlier message to ask a question, the system passes the quoted content to the AI as well, so the answer targets the message the customer is asking about. [User Guide](https://docs.maiagent.ai/build/choose-channel)
* **\[Shared] Message Webhook adds signature verification and automatic retry**: Message event notifications (Webhook) now include a signature and timestamp, so the receiver can verify that a notification truly came from MaiAgent and guard against replay; on temporary failures the system retries automatically. [User Guide](https://docs.maiagent.ai/org/developer)

### 🗂️ Full Updates (by Product Line)

<details>

<summary>Serve (15 items)</summary>

#### ✨ New Features

* **\[Serve] Conversations automatically return to AI when a human agent does not reply in time**: Conversation platforms gain a new "Auto-return to AI when a human does not reply in time" setting. After a human agent takes over, if no one replies to the visitor's message within the number of seconds you set, the system automatically hands the conversation back to the AI assistant and can send a custom notice at the same time. The timer starts the moment the agent first takes over, a scheduling failure no longer swallows the handoff notification, and messages that arrive at the instant of takeover are not missed. Suited to teams with limited support staff who worry that visitors are left waiting with no one attending. [User Guide](https://docs.maiagent.ai/build/conversation-settings/human-handoff)
* **\[Serve] Quoted (replied-to) message content is passed to the AI**: On platforms such as LINE that support quoted replies, when a customer quotes an earlier message to ask a question, the system passes the quoted content to the AI as well, so the answer targets the message the customer is asking about. [User Guide](https://docs.maiagent.ai/build/choose-channel)
* **\[Serve] Resolved conversations are reassigned to the original agent when the customer returns**: Conversation platform settings gain a new option: when a customer writes again within the time limit on a conversation marked as resolved, the system preferentially assigns it to the agent who handled it before, so the customer does not need to explain the background again. This option is off by default; turn it on in the conversation platform settings when needed. [User Guide](https://docs.maiagent.ai/build/conversation-settings/human-handoff)
* **\[Serve] LINE conversation platforms configure and verify the Webhook automatically**: When you create a LINE conversation platform, the system automatically sets the Webhook URL and verifies that the connection works, so you no longer need to paste the URL into the LINE console manually. [User Guide](https://docs.maiagent.ai/build/choose-channel)
* **\[Serve] Messenger／Instagram detect other subscribed apps and handle conversation-routing state**: When a Messenger or Instagram page also has other subscribed apps, the system detects them and recognizes the conversation-routing standby state, so it does not compete with other apps to reply. [User Guide](https://docs.maiagent.ai/build/choose-channel)
* **\[Serve] Speaking-member list and member count for LINE groups**: A new API returns the list of members who have spoken in a LINE group conversation and the group's total member count, helping support staff and external systems understand group participation. [User Guide](https://docs.maiagent.ai/build/choose-channel)
* **\[Serve] Videos and voice messages sent by human agents on LINE use native LINE messages**: When a human agent sends a video or voice message from a conversation platform, LINE users receive native video and voice messages that play directly inside LINE, instead of links. [User Guide](https://docs.maiagent.ai/build/choose-channel)
* **\[Serve] Agent can handle Telegram voice, video and audio files**: The AI assistant can now transcribe and understand voice messages and videos from Telegram, as well as audio and video sent as files; for videos, the audio track is extracted and transcribed, and the video length is reported. [User Guide](https://docs.maiagent.ai/build/choose-channel)
* **\[Serve] In group conversations, Agent sees messages that do not @ it, along with who spoke**: In group conversations on LINE, Telegram, Slack, Teams and LINE WORKS, the Agent now sees messages that do not @ it, along with the speakers, so it replies with the full discussion context in hand. [User Guide](https://docs.maiagent.ai/build/choose-channel)
* **\[Serve] Web Chat contact history can list conversations with no question yet**: The signed-in Web Chat contact history list can now also show conversations that have no messages yet, so Ask AI can reuse a conversation in which nothing has been asked. [User Guide](https://docs.maiagent.ai/serve/daily-operations/reply)
* **\[Serve] OpenAI gpt-live-1 voice billed per second**: Voice support adds billing for OpenAI gpt-live-1: the voice layer is billed per second, backend reasoning tokens are charged separately, and Credit usage records reflect actual cost more closely. [User Guide](https://docs.maiagent.ai/serve/voice)
* **\[Serve] Voice monitoring: delegation links, summary fields and dashboard metrics**: Voice monitoring is more complete: 1. Sessions show their delegation links; 2. The session list gains a summary field, with filtering and sorting; 3. The dashboard adds latency and delegation metrics; 4. Historical sessions and older Agents also get conversation turn counts; 5. The filter can select multiple assistants at once. [User Guide](https://docs.maiagent.ai/serve/voice)

#### 😊 User Experience and Interface Improvements

* **\[Serve] LINE stickers are now presented to the AI as text and images**: 1. LINE stickers are stored as "\[Sticker: keyword]", and attachment file names no longer contain the sticker ID; 2. When a sticker has no keyword, the model is shown the image instead, while stickers with keywords are still passed as text only; 3. Deep Research internal assistants and image limits no longer pick up LINE stickers by mistake.

#### 🛠️ Bug Fixes

* **\[Serve] Fixed issues on the LINE / Messenger / Instagram / Email / WhatsApp channels**: 1. Filled the gaps in attachment sending and receiving for human agents on LINE, Messenger, and Instagram; 2. Document attachments received by Email are now converted into files searchable in the conversation, as on other platforms; 3. Attachments uploaded from the WhatsApp support console are now uploaded to Meta before being sent, and AI-replied images are sent as images only, without the markdown source; 4. Messages recalled by LINE users or on Instagram, and messages deleted on Slack, no longer leave their original text for the AI to read; 5. LINE group mentions are now matched on name boundaries, so names that merely start the same no longer trigger a wrong reply; 6. Resent copies of the same message from LINE and Team+ no longer leave false failure records and alerts; 7. When a non-owner member belongs to multiple groups, the LINE monthly message quota and the bookmarked conversation platforms no longer return a 500 error; 8. Anonymous guest conversation data now includes the conversation platform ID.
* **\[Serve] Fixed conversation handling status and Copilot suggestions**: 1. Marking as read no longer writes old values back over "switch back to AI Agent" and assignment results; 2. For conversations outside AI auto-reply hours (including when all seven days are turned off), the handling progress no longer reports "AI is processing"; 3. Support Copilot suggestions are now generated from the conversation as it was at the moment of that visitor message, and drafts are no longer written back into AI memory.

</details>

<details>

<summary>Build (17 items)</summary>

#### ✨ New Features

* **\[Build] AI Gateway adds OpenAI-compatible audio endpoints**: AI Gateway adds OpenAI-compatible speech-to-text and text-to-speech endpoints. Existing applications that use OpenAI audio interfaces can connect by changing only the URL and key, and usage is deducted from Credits based on actual consumption. [User Guide](https://docs.maiagent.ai/build/ai-gateway)
* **\[Build] AI assistant list adds a "Creator" filter**: The AI assistant list can now be filtered by creator, so organizations with many assistants can quickly find those created by a specific member, which helps with inventory and handover. [User Guide](https://docs.maiagent.ai/build/agent-builder)
* **\[Build] New assistants enable only internal chat by default, with configurable defaults for web Widget and API access**: To reduce the risk of accidentally opening external channels: 1. Newly created AI assistants enable only internal chat by default, and you can choose whether to enable the web Widget and API access when creating one; 2. The organization console adds a setting that determines whether the conversation platforms and API access of new assistants are enabled by default, so the organization can set a unified policy. [User Guide](https://docs.maiagent.ai/build/agent-builder)
* **\[Build] Team canvas adds a routing entry node**: The team canvas adds a "Routing Entry" node. At the start of a team flow, it can examine the incoming message and then dispatch it to the right member Agent, making the routing logic of multi-Agent teams clearer. [User Guide](https://docs.maiagent.ai/build/agent-builder)
* **\[Build] Conversation records can be filtered by tool, and tool usage can be exported**: AgentOps tracing is more complete: 1. Conversation records can be filtered by the tools used, and tool usage can be exported; 2. Built-in tool calls are also shown in the tool execution records; 3. Step cards in team execution records show the interaction type and input/output summaries, making flows easier to trace. [User Guide](https://docs.maiagent.ai/build/agent-ops)
* **\[Build] MCP Apps: interactive card support**: Conversations on the Web channel now declare interactive card support to MCP servers, so MCP servers built with the official ext-apps SDK can display interactive cards directly in the conversation. [User Guide](https://docs.maiagent.ai/build/tools)
* **\[Build] Estimate Credit cost before a dataset evaluation**: AgentOps dataset evaluations can now estimate the Credits required before running, and an API is provided to retrieve the estimate, so large evaluations don't exceed the budget. [User Guide](https://docs.maiagent.ai/build/agent-ops)

#### 🚀 Core Performance Improvements

* **\[Build] Lower document parsing cost: Vision Parser resume and reuse**: Vision Parser parses large documents more cheaply: 1. When output is truncated, pages that are already complete are kept and only the remaining pages are resent; 2. Checkpoints are created per page, so a rerun after an interruption only fills in unfinished pages; 3. When a file with identical content is re-uploaded, the existing parsing result is reused without resending to Bedrock, the cache connection has a timeout guard, and parsing proceeds as usual if Redis hangs; 4. Vision Parser can now run through Azure OpenAI (GPT-6.1 Sol).
* **\[Build] MCP tool list caching**: The MCP tool list (tools/list) for the same identity is now cached and sorted consistently, so repeated loads no longer request it from the MCP server every time, and the catalog order is more stable.
* **\[Build] Lower evaluation judge call cost**: Judge model calls in dataset evaluations are cheaper: the wallet lock is now acquired later when deducting Credits, reducing peak queuing in parallel evaluations; judge calls also no longer write to the prompt cache.

#### 😊 User Experience and Interface Improvements

* **\[Build] Clearer model selection and lists**: 1. All model dropdowns in the admin console can now distinguish models with the same name; 2. The card descriptions of the 8 official retail Agents in the Agent Marketplace now contain concrete content; 3. The default configuration automatically filled in for voice "STT + LLM + TTS" is now Taiwanese Mandarin (previously English); 4. Call titles now use the first sentence spoken, and replies read aloud by the assistant are labeled as voice transcripts.
* **\[Build] Clearer MCP connection error reasons**: When configuring MCP tools: 1. A failed "Get Tools" now states the reason (certificate, 401, DNS, connection, timeout) instead of always showing "No tools available"; 2. When the server rejects the request (401, blocked redirect, 5xx), the reason is shown immediately instead of waiting 30 seconds to show a timeout; 3. Error events no longer carry URLs and credentials into monitoring; 4. Query parameters of API tools mask only credential-like values, and ordinary data is passed to the AI as is.

#### 🔐 Security and Governance Enhancements

* **\[Build] Tighter permissions for Hooks, teams, presentations, and templates**: 1. Only the owner or creator can modify and delete Hooks and teams, and Hook execution records include only AI assistants the user can see; 2. Public share links for AI presentations can be disabled, generating and disabling them is limited to the creator or organization owner, and the AI edit entry point is shown according to permissions; 3. Assistants created by applying a template are no longer opened directly to the default role (the whole company); 4. Text-to-SQL on the uploaded-file shared library now has a table allowlist, blocking system catalogs and dangerous functions; 5. Test set permissions now cover all reads, and generated test questions can only use knowledge bases the member can read.

#### 🛠️ Bug Fixes

* **\[Build] Fixed issues with AI assistant tools, skills, and image understanding**: 1. When a file write is left unfinished, the model is reminded to finish and continue; 2. Rewriting the same file in the same turn is no longer blocked as a "duplicate call"; 3. When tool search is on, bound knowledge base retrieval tools are loaded directly, and the mandatory-reference instruction is added only when the knowledge base was actually read; 4. When knowledge base retrieval results exceed the tool output limit, complete passages are kept by rank and only low-scoring nodes are dropped; 5. The attachment parsing tool description recommends only parsers enabled in this environment; 6. When an image is re-examined with a tool, the model receives the image and no longer fabricates receipt contents; 7. When generation fails after a video confirmation card is submitted, the failure notice appears under the "Video task submitted" reply; 8. Word-to-PDF conversion preserves DFKai-SB and PMingLiU fonts, and charts that specify Windows fonts no longer show Chinese characters as boxes.
* **\[Build] Fixed document parsing (OCR) issues**: 1. The OCR parser now applies PDF page rotation, so documents that are displayed rotated after a landscape scan are no longer recognized as 0 characters; 2. PDFs whose fonts lack character code mappings now use full-page OCR instead of garbled text; 3. Large knowledge base files are no longer killed midway through parsing; 4. Qwen's provider-default thinking is turned off during parsing and image description, and a clear message is shown when content moderation blocks a request; 5. Uploading a file that does not exist returns 400, the knowledge base batch upload description no longer states a false 50-file limit, and sending a message attachment with a missing id or an invalid UUID returns 400; 6. FAQ Excel import skips sample rows and detects duplicates by question text; 7. When a crawl returns 0 pages and no error can be found, "Unable to retrieve error details" is shown, and metering fields are now recorded; 8. When an attachment has no content after parsing or indexing fails, the chat window is notified, so the attachment no longer shows as processing indefinitely; 9. When object storage uses a TWCA certificate, image and batch Q\&A file downloads apply the relaxed SSL domain setting, so images can be sent to the model normally.
* **\[Build] Fixed voice and scheduling issues**: 1. Evaluation and scheduled jobs no longer exceed the queue's 30-minute reporting limit and cause the entire worker to restart, which also resolves knowledge base processing getting stuck; 2. When a skipped shared-tool authorization has expired, the schedule is recorded as "Partially completed" and the creator is notified, and the Agent also tells the user that reauthorization is required.
* **\[Build] Fixed Text to SQL and database selection issues**: 1. Date columns in Excel files uploaded to a database are now created as date types, so "up to a given date" queries no longer miss the last day; 2. Before execution, table name case, database names, PostgreSQL backticks, and string colons are corrected automatically, and column errors list the actual columns with a cap on the number of columns shown; 3. Shared-database table name truncation no longer causes mismatches, and the reconciliation command can restore orphaned tables and handle tables stuck for more than 24 hours or that failed to delete; 4. When the AI assistant's "Select database" list had more than 20 databases, it was incomplete, so the list is now unpaginated.

</details>

<details>

<summary>Shared (21 items)</summary>

#### ✨ New Features

* **\[Shared] Batch role permission settings and effective permissions for members**: Organization management adds two permission governance capabilities: 1. Apply permission settings to multiple MaiAgent roles at once instead of editing them one by one; 2. View the permissions actually in effect for a single member and see which role each permission comes from. The permission screens were also tidied up: role permission names are now fully in Chinese and consistent with the screen, the role permission list is sorted by display name, missing translations for Hook and conversation platform analytics permissions were added, and the quota self-check screen now shows the reminder threshold. [User Guide](https://docs.maiagent.ai/org/roles)
* **\[Shared] Message Webhook adds signature verification and automatic retry**: Message event notifications (Webhooks) now include a signature and timestamp, so receivers can verify that a notification really comes from MaiAgent and guard against replay; the system automatically resends when a temporary failure occurs. [User Guide](https://docs.maiagent.ai/org/developer)
* **\[Shared] Usage statistics scope can be set per role**: The visible scope of usage statistics can now be set per MaiAgent role: choose "Entire organization" or "Authorized AI assistants only", with the entire organization as the default. Organizations that need to limit certain roles to seeing only the usage of the assistants they are responsible for can adjust this directly. [User Guide](https://docs.maiagent.ai/org/roles)
* **\[Shared] Organization reminder announcements and one-time notifications**: Announcements add an organization reminder method: administrators can send reminders to organization members, and each member sees a one-time notification that no longer pops up again once acknowledged. Suitable for publishing maintenance notices or policy updates. [User Guide](https://docs.maiagent.ai/org/overview)
* **\[Shared] On-premises deployment adds Breeze-ASR speech recognition**: On-premises LocalAI adds the Breeze-ASR-26 and Breeze-ASR-25 speech recognition models, with LocalAI speech recognition settings pre-created (disabled by default) at three entry points; they can be used once enabled, and real-time pricing is supported. The default silence detection time for real-time speech recognition was changed from 100 ms to 500 ms, so speech is less likely to be cut off at pauses. [User Guide](https://docs.maiagent.ai/serve/voice)
* **\[Shared] Organization source IP allowlist**: Organizations can set an allowlist of source IPs permitted to sign in and use the platform: once set, features and APIs that require sign-in accept only sources on the allowlist, and the organization settings API can also read and update the allowlist, strengthening enterprise network boundary control. [User Guide](https://docs.maiagent.ai/org/overview)
* **\[Shared] Cohere Embed 5 Pro / Fast embedding models**: Adds the Cohere Embed 5 Pro and Fast embedding models, which can be selected for knowledge bases to strike the right balance between accuracy and speed. [User Guide](https://docs.maiagent.ai/build/knowledge-base)
* **\[Shared] Several new models added to the model catalog**: Selectable models are updated: 1. Added Claude Haiku 5.5 (AWS Bedrock Japan region); 2. Enabled Qwen3.8 27B, which will take over from Qwen2.5 VL 72B, and assistants using the old model are migrated automatically; 3. Added GPT-6.1 Sol via direct OpenAI, ordered before the Azure version; 4. Enabled Muse Glimmer 30B; 5. Added Qwen3.6 27B (the same model shipped on-premises, fixed to DeepInfra fp8, with no China-based providers); 6. Enabled DeepSeek V4.1 Flash on Azure, with GPT-6 Luna announced as its successor in December 2026. [User Guide](https://docs.maiagent.ai/build/ai-gateway)
* **\[Shared] Model retirement and succession: old models retired, assistants migrated automatically**: The model catalog is being refreshed and tidied: 1. A second batch of 7 low-usage models (three Watsonx models, Mistral Medium 3.5, Qwen3.5 9B and 397B, Nova 2 Lite) is retired, and assistants in use are migrated automatically to the successor models, while o3 stays scheduled to retire on October 20, 2026; 2. The direct OpenAI entries for GPT-5.6, GPT-5.5, and GPT-5.4 mini are retired in favor of the Azure versions; 3. GPT-OSS 120B now defaults to Azure with the model ID unchanged, and old keys are not carried over when the provider changes; 4. Gemini 2.5 Flash is succeeded by 3.5 Flash-Lite after retirement; 5. System background tasks now pick a Gemini Flash by environment version (3.8 Flash in production); 6. The model menu now uses a fixed order: tier first, then newest to oldest within a tier (Claude in the order Fable, Opus, Sonnet, Haiku). [User Guide](https://docs.maiagent.ai/build/ai-gateway)
* **\[Shared] Centralized model settings for system tasks**: The admin console groups the models used by "system tasks" into a single settings group, and the AI memory task now uses an explicitly specified model instead of automatically picking the latest Gemini Flash, making cost and behavior more predictable. [User Guide](https://docs.maiagent.ai/build/ai-gateway)

#### 🚀 Core Performance Improvements

* **\[Shared] Faster AI assistant reply startup and wrap-up**: Several steps in the reply flow now wait less: 1. The embedding client uses a shared connection pool, removing one Bedrock connection handshake before the first model call; 2. When knowledge base results were already ranked by the assistant's own reranker, they are no longer reranked a second time; 3. The AI assistant list no longer deduplicates across all fields, saving about 0.15 seconds per request; 4. Customer service conversation page counts no longer time out for non-owner members of large organizations; 5. The last message of a non-streaming conversation no longer goes backwards; 6. The schema on the API documentation page is now cached, so it opens faster.

#### 😊 User Experience and Interface Improvements

* **\[Shared] More complete activity logs**: 1. The activity log records additions and removals in an API key's "enabled models" list; 2. The exported "API Path" column is no longer blank; 3. Organization setting change history is now complete, and "update" events with no changes are no longer recorded; 4. Conversation records from scheduled runs show credit consumption and processing time; 5. Conversation record exports now include the node IDs of cited sources.
* **\[Shared] Minor improvements to chat platform and schedule settings**: 1. Partially updating a chat platform no longer requires the full theme, and the locale zh-tw is saved as zh-hant; 2. Schedule timeout errors are shown in the user's language; 3. Organization data now includes the minimum schedule interval, and the admin schedule form uses it to show the lower limit; 4. The Traditional Chinese wording for "identity" is unified to the standard form 「身分」.

#### 🔐 Security and Governance Enhancements

* **\[Shared] Stronger access isolation for conversations and sharing**: Conversation access permissions are stricter: 1. A colleague's MaiGPT conversations can no longer be shared; sharing is limited to conversations you can read; 2. Non-owners see only their own MaiGPT conversations (list, details, and counts); 3. On the personal chat surface, members cannot open a colleague's conversation; 4. Unauthorized callers can no longer read Copilot suggestion drafts for customer service, and drafts in existing shares are filtered out; 5. Choosing the Deep Research output format is limited to conversation members; 6. Thinking-setting updates respond according to the branch that actually admitted the caller.
* **\[Shared] Cross-organization access protection: attachments, schedules, continued chats, and Web Chat**: Closes cross-organization access gaps: 1. Continuing a chat and creating a schedule accept only Agents and conversations within the caller's organization; 2. Schedules accept only Agents and chat platforms from the caller's organization; 3. Attachments can be bound only to attachments of the current conversation, conversations in channels disabled by another organization return 404, and non-object request bodies return 400; 4. Web Chat message sending and history are isolated by chat platform, and a data-fix command is provided; 5. Public registration is bound to the invited recipient.
* **\[Shared] Stronger sign-in and authentication**: Sign-in and invitation flows are more secure: 1. LDAP members can no longer bypass LDAP sign-in with "Forgot password," and an owner saving third-party sign-in settings no longer switches "Enable sign-in" back on; 2. The 30-day OAuth refresh token lifetime now takes effect, and expired tokens are cleaned up daily; 3. The OAuth authorization sign-in page supports signing in with an organization account (SSO); 4. Regular members can no longer create or copy organization invitations, and the owner group can be invited only by owners; 5. The API key requests-per-minute limit now also applies on endpoints with their own rate limits; 6. A 401 for failed API key authentication now points to Api-Key, and an invalid key returns "Invalid API key"; 7. Authentication source names for all four third-party types follow the same rule (lowercase letters and digits plus hyphens, up to 31 characters), and existing names are unaffected; 8. When the registration page URL is not configured, invitations are marked as email delivery failed instead of silently staying pending.
* **\[Shared] Logs no longer store full keys, and API request logs are kept for 30 days**: 1. Partner API and Webhook logs no longer store full keys, and message Webhook headers are now write-only; 2. Partner and regular API request logs are kept for only 30 days and are cleared automatically every night.
* **\[Shared] Stricter Credits balance checks**: Prevents credits from being deducted into negative numbers: 1. Before a conversation reply, the wallet must cover the estimated credits for one reply; 2. Before parsing knowledge base files, credits are estimated from page count and parser, and parsing that cannot be afforded is blocked; 3. API key and JWT callers go through the same Credits check as browser sign-in, and AI Gateway checks personal and role quotas; 4. Once a balance has been judged insufficient, an error in auto top-up or balance lookup still returns 402 and blocks the request.

#### 🛠️ Bug Fixes

* **\[Shared] Fixed interrupted AI replies, resends, and streaming errors**: 1. When a stalled Bedrock stream is retried, it is replayed as a clean message that langchain can parse; 2. Redacted OpenAI reasoning content no longer breaks Agent streams; 3. Agent model streams that are closed right at the start are resent automatically; 4. Rate-limit and transient server errors reported by Azure/OpenAI mid-stream are now retried; 5. A reply cut off by a worker restart or shutdown is shown as a visible, unbilled failure; 6. Fixed the error tail after stopping a reply on-premises; 7. Hooks now flush previously held text before tool items and at the end of the stream.
* **\[Shared] Fixed inaccurate Credits billing**: 1. OpenAI direct-connection prompt cache write tokens are now billed as LLM Cache Creation Tokens; 2. Replies with no output tokens now also count the usage that was recorded; 3. Unrecorded embeddings are no longer charged; 4. A completed video reply no longer attributes the next round's tool costs to the round that submitted the video; 5. Vision parsing deducts credits every time, so re-parsing, interrupted reruns, and new file library versions are no longer missed; 6. When two evaluation settlements run at the same time, the fixed fee continues from the number of questions already billed by the previous one, and judge calls are counted only once; 7. Platform-key usage by system features (automatic evaluation, AI drawing coordination, web fetch summaries, document image descriptions) records cost only, at 0 credits.
* **\[Shared] Fixed assistant, organization, and AI Gateway settings issues**: 1. After a fallback group is deleted, bound Agents reply and bill using their own models; 2. Bulk-importing member Metadata now merges, so attributes not listed in the file and personal preferences are no longer cleared; 3. The PII masking Hook template now masks all cards when two cards appear back to back, and on-prem licensed deployments now include 5 system Hook templates; 4. On-prem deployments now include the missing database examples, and "Add database" can select SynxDB; 5. The Monday connector icon now points to the public CDN after the domain change; 6. In-app notification links now use relative paths, so clicking "Connector enabled" on on-prem and white-label deployments no longer leads to the SaaS domain; 7. Non-streaming replies no longer record output speed, Agent-mode replies record output tokens per second, and the model access page no longer shows hundreds of millions of tokens per second.

</details>

### 🗂️ Full updates (by product line)

<details>

<summary>Work (18 items)</summary>

#### ✨ New features

* **\[Work] Outlook email file (.eml/.msg) parsing**: Knowledge bases and conversation attachments can now process Outlook email files directly: 1. After you upload an .eml or .msg file, the system parses the message body and sender information and lists its attachments; 2. It also extracts content from embedded document attachments, such as quotes, PDFs, and Excel files, so AI can answer questions about their details; 3. Email attachments forwarded through Outlook are recursively expanded; 4. S/MIME-signed or encrypted emails are handled according to their status, and encrypted content that cannot be read is marked; 5. Deleting an email file also removes the attachments extracted from it; 6. TNEF (winmail.dat) files are detected and marked as present without being expanded. Useful for teams organizing customer emails and quote exchanges in a knowledge base or dropping them into a conversation to ask questions. [User guide](https://docs.maiagent.ai/build/knowledge-base)
* **\[Work] Work Agent file workspaces: mount File Library folders on custom and official Agents**: File Library folders can be mounted on Work Agents as workspaces: 1. Custom Agents can have a bound workspace and can allow users to mount their own; 2. Official Agents allow users to mount their own folders; 3. Mounting now supports any nested subfolder, and folder file counts include subfolders. Files modified by an Agent in a workspace can be saved back to File Library. [User guide](https://docs.maiagent.ai/work/file-library)
* **\[Work] Bulk deletion of your conversations**: Select and delete multiple conversations you own from the conversation list in one action, making it faster to clean up a large backlog. [User guide](https://docs.maiagent.ai/work/maigpt)
* **\[Work] File Library storage usage and a dedicated Credit category**: Two additions make File Library usage more transparent: 1. Organization overview can report File Library storage usage; 2. Credit usage records include a dedicated service category for File Library, with derived content storage accounted for separately by source, making storage and Credit spending easier to track. [User guide](https://docs.maiagent.ai/work/file-library)
* **\[Work] Meeting context for speech recognition in meeting records**: Speech recognition in meeting records now uses meeting context: 1. Calendar invitation content is applied before a meeting is created, providing recognition cues such as the meeting room name and participants; 2. Context is built and sent to the Agent before live recognition starts; 3. Offline recognition also receives context when importing files or retranscribing; 4. A feature flag lets administrators control the capability. This improves recognition of specialized terms and people's names. [User guide](https://docs.maiagent.ai/work/meeting-records)
* **\[Work] Separate AI memory pricing and transparent usage**: AI memory Credit usage is now clearer: 1. Memory extraction is billed separately against the quota of the member being remembered; 2. Credit usage records include an AI memory service category; 3. Response input details record the number of tokens used by AI memory; 4. Summaries filter out personal information, and malformed output is not charged. [User guide](https://docs.maiagent.ai/work/maigpt)
* **\[Work] Structured organization of AI memory**: AI memory is now better organized: 1. Memories are stored by category, and new information merges with existing memories to avoid duplicates; 2. User facts are extracted together after a conversation has been quiet for a while; 3. Memory topics appear as titles in the user's language for all organizations. [User guide](https://docs.maiagent.ai/work/maigpt)

#### 🚀 Core performance improvements

* **\[Work] Faster, less expensive memory background tasks**: Memory background operations use fewer resources: 1. Asking the assistant to remember something no longer requires an extra round trip to the main model, and memory included in each turn is capped at 600 tokens; 2. Background calls use system task models without thinking, follow organization rules, and fall back to the main model only on failure; 3. Long-conversation consolidation uses system task models, with cost guardrails configurable in Admin; 4. If a system model fails, the task retries once using the next system task model; 5. A model response stating that there are no facts no longer triggers a fallback rerun, reducing excessive fallback rates.
* **\[Work] Higher prompt cache hit rates for long conversations**: Prompt caching for OpenAI, Azure, and Claude is more stable: 1. Long-term memory is moved out of the cached prefix so Claude no longer rewrites the entire cache on every conversation turn; 2. Short-term memory compression now reduces usage to 60% of the budget and compresses in batches when the protected portion itself is too large, keeping the start of history stable across turns; 3. The current time is moved to the end of the message, and each user message carries its send time, allowing long conversation histories to hit the cache; 4. Cache routing now uses the assistant as the key, allowing new conversations to hit the cache too (GPT-5.6 previously missed every time).
* **\[Work] More concise Canvas Agent prompts**: Tool results are no longer repeated in the Canvas Agent prompt after handoff, shortening prompts and reducing model costs.

#### 🔐 Security and governance enhancements

* **\[Work] GPT image generation and editing no longer retain content at OpenAI**: Content is no longer stored at OpenAI when generating or editing images with GPT. Edits now send the locally stored previous image, reducing the risk of image data remaining with a third party.

#### 🛠️ Bug fixes

* **\[Work] AI memory truncation and toggle issues fixed**: 1. When older tool results are truncated, knowledge base content retains 100 characters and any IDs removed by truncation; only the copy written to history is truncated; 2. IDs are also preserved when JSON printed by Code Interpreter is truncated; 3. When AI memory is off, the assistant no longer promises to remember things; 4. Fixed a regression from these changes: with memory off, the assistant no longer responds to ordinary requests with “I cannot assist”; 5. When a message is recalled during a response, memory still retains the attachment IDs of photos from the same turn.
* **\[Work] File Library downloads and conversation attachment issues fixed**: 1. Downloads of File Library conversation attachments and generated files no longer return a 500 error when the file exists; 2. Downloads retain the display filename instead of using a 32-character hash; 3. Regular members can open File Library conversation attachments without loading failures (listing, downloading, and early deletion all work); 4. Lists and downloads of conversation-generated files no longer return 500 errors after the assistant is deleted; 5. MaiGPT conversation attachment lists include the workspace source (Workspace › Folder).
* **\[Work] Fixed file library version history and search**: 1. Version history messages follow the interface language; 2. Search results include creation time and the last updater; 3. File counts can include subfolders; 4. Version history identifies who approved Agent deletions, renames, moves, and edits, along with the source conversation. Source conversations for generated files show their display names.
* **\[Work] Fixed confirmation and write-back for Agent file operations**: 1. Agent file edits and renames succeed on the first confirmation; 2. Office extraction results explicitly instruct the Agent to write changes back to the file library; 3. When the Agent modifies a sandbox copy of a workspace file, it receives a reminder to write it back; 4. Files generated by code are saved to the file library's temporary storage as soon as they are written, without opening the Output panel; 5. Generated images and videos are recorded directly without queueing.
* **\[Work] Fixed file library images, folders, and notifications**: 1. Uploaded file library images are immediately marked as ready to load instead of showing a knowledge base parsing failure; 2. Empty folders and file workspaces can be deleted; 3. Deletion approval notifications include a workspace link; 4. The Output panel in Work conversations reads from the file library and identifies the owner from the source when the contact's user is empty; 5. The current-conversation file tool omits image lists for non-Agent queries, all tools block removed files, and Canvas lists only the latest version.
* **\[Work] Fixed Canvas and presentation assistant responses**: 1. After calling a nonexistent tool, the Canvas agent is prompted to write directly to Canvas instead of switching to location search; 2. A route map no longer appears as raw JSON after a map was drawn in the previous turn; 3. AI presentation responses include the current turn's image and layout status, so unconfirmed images are no longer described as already inserted into the presentation.
* **\[Work] Fixed voice and meeting record issues**: The meeting transcription context field now has a database default, preventing meeting creation failures during rolling deployments. Meetings now start with a context prepared in advance.

</details>

## v2026.10.09 <a href="#v2026-10-09" id="v2026-10-09"></a>

***

### 📌 Release Highlights

* **\[Serve] Fix conversation ownership across multiple Web Chats**: When multiple Web Chats are embedded on the same page and use the same user identifier (Source ID), creating a conversation with a contact from another conversation platform now uses the contact for the same user on the current platform instead, so the new conversation is created successfully and assigned correctly. This fix applies when both platforms belong to the same organization, a matching contact already exists on the current platform, and Source ID access verification is disabled.

### 🗂️ Full Updates (by Product Line)

<details>

<summary>Serve (1 item)</summary>

#### 🛠️ Bug Fixes

* **\[Serve] Fix conversation ownership across multiple Web Chats**: When multiple Web Chats are embedded on the same page and use the same user identifier (Source ID), creating a conversation with a contact from another conversation platform now uses the contact for the same user on the current platform instead, so the new conversation is created successfully and assigned correctly. This fix applies when both platforms belong to the same organization, a matching contact already exists on the current platform, and Source ID access verification is disabled.

</details>

## v2026.10.07 <a href="#v2026-10-07" id="v2026-10-07"></a>

***

### 📌 Release Highlights

* **\[Build] Restored the Quick Reply creation option**: On the type selection page for adding an Agent UI, Quick Reply now appears alongside Flex Message and Quick Menu. You can open the existing editor directly to configure interactive reply buttons.
* **\[Build] Improved knowledge base retrieval after follow-up questions**: When you use an AI assistant in general mode and answer its follow-up question with a short phrase, such as a location, the system uses the conversation text to formulate the knowledge base search query. This reduces the chance of retrieving irrelevant content by overlooking the original question.
* **\[Shared] AI memory management restricted to your own memories**: “AI Assistant → Advanced Settings → AI Memory” now lists and manages only the signed-in user's own memories. The same restriction applies to organization owners. Viewing, editing, deleting, and bulk deleting memories are all protected, with no additional configuration required.

### 🗂️ Full Updates (by Product Line)

<details>

<summary>Build (2 items)</summary>

#### 🛠️ Bug Fixes

* **\[Build] Restored the Quick Reply creation option**: On the type selection page for adding an Agent UI, Quick Reply now appears alongside Flex Message and Quick Menu. You can open the existing editor directly to configure interactive reply buttons.
* **\[Build] Improved knowledge base retrieval after follow-up questions**: When you use an AI assistant in general mode and answer its follow-up question with a short phrase, such as a location, the system uses the conversation text to formulate the knowledge base search query. This reduces the chance of retrieving irrelevant content by overlooking the original question.

</details>

<details>

<summary>Shared (1 item)</summary>

#### 🔐 Security and Governance Enhancements

* **\[Shared] AI memory management restricted to your own memories**: “AI Assistant → Advanced Settings → AI Memory” now lists and manages only the signed-in user's own memories. The same restriction applies to organization owners. Viewing, editing, deleting, and bulk deleting memories are all protected, with no additional configuration required.

</details>

## v2026.10.06 <a href="#v2026-10-06" id="v2026-10-06"></a>

***

### 🛠️ Bug Fixes <a href="#v2026-10-06-bug-fixes" id="v2026-10-06-bug-fixes"></a>

* **Fixed failed Bedrock response retries**: When a response stalls or times out while using an Amazon Bedrock model, the system now correctly resumes the response if an automatic retry has already returned an answer. This prevents errors in processing retry content from interrupting the response or affecting the next conversation turn.
* **Recovery after Canvas generation failures**: When creating charts or dashboards in Canvas mode, if the AI mistakenly uses a tool that cannot be executed, the system guides it back to the Canvas output flow. This reduces cases where only text is returned without generating a Canvas, as well as unnecessary point consumption from mistakenly using location search. This fix also covers the Canvas flow in assistant teams.
* **Fixed missing text at the end of Web Chat responses**: AI assistants with response masking or checking enabled now complete the checks and send the remaining text before finishing a response, preventing the ending from being omitted in Web Chat. Explanations preceding tool operations are also displayed in order, and existing content protection still applies.
* **Fixed the customer service entry disappearing after disabling transfers**: When an administrator edits a role that already has full access to customer service conversation pages, clearing “Transfer others' conversations” and saving now preserves the customer service conversation entry and existing page permissions. Selecting the parent customer service conversation permission also no longer automatically enables the transfer permission; it must be selected separately.

## v2026.10.05 <a href="#v2026-10-05" id="v2026-10-05"></a>

***

### ✨ New Features <a href="#v2026-10-05-new-features" id="v2026-10-05-new-features"></a>

* **Minimum turns for multi-turn evaluations**: When creating a multi-turn conversation evaluation, you can specify the minimum number of turns required before it can end upon reaching its goal, preventing tests from stopping too early. The results page also shows this setting.
* **Wait for user replies in assistant teams**: Automatic and handoff connections in assistant teams can wait for a user reply before continuing, with optional reply choices. This supports workflows that need to confirm information step by step.
* **Default Credits allowance for new members**: Organizations can set the Credits allowance, expiration, and reset cycle for new members, and override the allowance when sending an invitation. When the settings dialog is reopened, previously saved results will not overwrite the new form.
* **Gemini real-time voice model**: Voice assistants now offer the Gemini 3.8 Live model for administrators to configure real-time voice interactions. Existing voice settings will not switch automatically.
* **New available model options**: The model menu now includes Claude Sonnet 5.5 (AWS Bedrock) and GPT-6.1 Sol (Azure). Administrators can select them for AI assistants and MaiGPT according to their organization's authorization.
* **Find files in the current conversation**: You can ask the assistant to list, filter, or reopen files in the current conversation, including uploaded images and assistant outputs, making it easier to retrieve previously used content across channels.
* **Cross-conversation AI memory for cloud organizations**: AI memory is now available to cloud organizations, allowing assistants to refer to saved preferences and facts in members' subsequent conversations. Memory is separated by organization and member; personal memory is not saved for visitors who are not logged in.

### 🚀 Core Performance Improvements <a href="#v2026-10-05-performance" id="v2026-10-05-performance"></a>

* **Improved loading of large knowledge base menus**: The knowledge base menu in conversations loads names and file counts first, fetching files and FAQ only when an individual knowledge base is expanded. Large collections of knowledge bases no longer load all content at once.
* **Improved preparation for knowledge base follow-up questions**: Knowledge base Q\&A skips unnecessary question rewriting. When rewriting is needed, it uses the platform's configured lightweight model and checks the result, reducing extra waiting before an answer begins.
* **Reduced waiting for concurrent usage deductions**: When multiple usage charges are settled simultaneously within an organization, price lookups, usage logging, and notifications no longer hold the wallet for extended periods, reducing contention between Credits deductions.
* **Improved conversation history loading**: When loading conversation history with citations and files, repeated retrieval of document content and download links is reduced, preserving the existing response content while reducing waiting.
* **Improved assistant memory initialization**: When an assistant starts answering, the auxiliary model used to organize memory is now prepared only when needed, reducing extra initialization before the first reply.

### 😊 User Experience and Interface Improvements <a href="#v2026-10-05-user-experience" id="v2026-10-05-user-experience"></a>

* **Role permission explanations and assistant list updates**: Role settings clarify the difference between AI assistant management permissions and conversation platform usage permissions to help administrators assign them correctly. The available assistant list is rechecked when members switch conversations.
* **Evaluation cost traceability**: Evaluations are billed for actual usage, including the models and knowledge retrieval used to answer, with costs aggregated per evaluation. Credits usage records can open the corresponding results directly; members without evaluation permissions see a permission notice.
* **Improved provider name display**: AI Gateway's credential menu for custom models and provider overview now show recognizable provider names, avoiding internal codes or blank names.
* **Customer service assignment and send reminders**: 1. Assignment and transfer lists show customer service agents' online status; 2. Unassigned conversations can be assigned directly to an agent; 3. Sending a message before taking over a conversation prompts a reminder, with options to take over and send, send directly, or cancel.
* **Meeting recording and upload checks**: Before live recording starts, microphone availability is checked, with prompts for permission or device issues. Offline uploads also check audio file content, blocking non-audio files whose extensions have merely been changed.
* **Automatic system announcement updates**: While the admin page is open, unread announcements refresh every 60 seconds. Notices automatically reflect the latest status when model services recover or announcements are cleared, without requiring a page refresh.
* **Meta channel setup guidance**: When creating a Messenger, Instagram, or WhatsApp conversation platform, the interface reads the current deployment's connection settings. Channels with incomplete setup remain visible and show notices about missing settings.
* **Duplicate assistant name notices**: When creating, copying, or renaming an AI assistant, duplicate names within the same organization are flagged below the name field, making it easy to correct the name before submitting again.
* **File deletion activity records**: Knowledge base file deletion events in activity logs retain the filename at the time of deletion, allowing administrators to identify the affected file after it has been removed. Mobile detail views now use a single-column layout.
* **Schedule quotas and completion status**: The schedule page shows the number of active schedules, the organization's limit, and the remaining quota. Schedules that reach their execution limit are marked completed, distinguishing them from schedules that still consume quota.
* **Conversation platform description field**: Conversation platform settings now include a description field where administrators can explain the assistant's purpose, giving organization-built assistants a clear introduction on marketplace cards.
* **Assistant team information masking explanations**: Node settings clarify attachment and paraphrasing rules when original user input is disabled, helping administrators understand what information each node can actually read.
* **Default reasoning effort behavior explained**: The default option in model reasoning settings explains what happens when no effort level is specified, and higher reasoning levels show a notice that waiting time may increase.
* **Login error notices within the form**: When an account or password is incorrect, the notice appears directly in the login form, with a forgot-password link to help users correct their input or reset their password.
* **Time awareness enabled by default for new conversation platforms**: New conversation platforms enable time-aware replies by default, allowing assistants to answer based on the current time and time zone. Existing platform settings retain their values, and the current time is also shown when editing.
* **Corrected file workspace operation descriptions**: The workspace selector now explains that assistants can write and perform file operations according to user permissions, removing wording that could misleadingly suggest the workspace is entirely read-only.
* **Provider deletion impact notices**: Before deleting an AI Gateway provider, the confirmation dialog shows how many models still use its credentials, helping administrators understand the impact first.
* **Consistent interactions in the Work workspace**: 1. Search includes a clear action, and dangerous actions such as deletion explain their consequences first; 2. Form submission displays errors together and focuses the first invalid field; 3. Conversations and groups support double-click renaming, with the latest result taking precedence during rapid successive renames; 4. Failure notices remain visible longer, and progress notices stay until the operation finishes.
* **File output cards in Web Chat**: When an assistant or a delegated sub-assistant generates files, Web Chat displays file cards so visitors can view and download the content from each output.
* **Adjusted greeting feedback actions**: Web Chat greetings retain the copy action and hide thumbs-up and thumbs-down buttons, which are excluded from satisfaction statistics, focusing feedback on actual answers.
* **Localized Hook template display**: When selecting a Hook template, its name, type, and description appear in the current interface language, helping administrators using Traditional Chinese understand its purpose.
* **Automatic reassignment when an agent pauses new assignments**: When a customer service agent pauses new assignments, automatically assigned conversations that they have not yet replied to are reassigned to other available agents, or returned to the queue if no one is available. Conversations already replied to or manually assigned keep their original assignment.
* **Improved Reply Formatting in Messaging Apps**: Assistant replies in LINE, Messenger, Instagram, and WhatsApp are converted to plain text so formatting symbols for bold text, headings, and other styles do not appear directly. Content with large numbers of repeated symbols also no longer stalls during format conversion.
* **Combining Consecutive Messages in Messaging Apps**: When visitors send multiple images in succession or send an image followed by text, the system waits briefly and combines them into a single response. Messages in the same conversation are processed in order to prevent simultaneous replies across multiple turns.
* **Trimming Older Tool Results for New Assistants**: Newly created assistants trim older tool results based on content size by default, reducing the space that historical tool output occupies in long conversations. Existing assistant settings retain their current values.
* **Automatic Replacement of Retired Qwen Models**: After Qwen3.6 Max is retired, the daily retirement job switches assistants using it to Qwen3.6 Plus and adjusts output length limits to avoid further calls to the disabled model.
* **Gemma Service Provider Change**: Gemma 4 31B is now provided through AWS Bedrock. The former OpenRouter version is registered for replacement, and the retirement job migrates assistants to the new provider so the old provider is not made available again after deployment.
* **Notifications for Expired Shared MCP Authorization**: When authorization for a shared MCP connector expires, the system skips that service and notifies organization members who can manage tools. In-app announcements and email guide them through reauthorization. Notifications are not repeated during the same authorization outage.

### 🔐 Security and Governance Enhancements <a href="#v2026-10-05-security" id="v2026-10-05-security"></a>

* **Sensitive Tool Information Masking Settings**: When editing a tool, you can specify sensitive headers to mask and view the platform's default masking entries, helping prevent authentication information from appearing in tool execution logs.
* **Model Origin Filters and Visibility Settings**: The model menu provides origin labels and filters, and organizations can centrally hide Chinese models. Model selection for web scraping also follows this restriction.
* **Permission to Transfer Conversations Served by Others**: Only members with permission to transfer other agents' conversations can transfer or take over a conversation that someone else is serving, preventing unauthorized assignments from interrupting ongoing support.
* **Stronger Meta Message Source Verification**: Incoming Messenger and Instagram messages undergo source verification. Deployments without the required verification information reject messages, preventing unverified requests from posing as channel messages.
* **Login Restrictions for Disabled Authentication Sources**: When an administrator disables login for an authentication source, Keycloak, EIP, SAML, LDAP, and linked third-party logins all honor the restriction, preventing login access through a disabled source.
* **Access Logs for Expired API Keys**: Unauthorized requests using retired API keys are recorded in the corresponding organization's API access logs, helping administrators trace connections still using old credentials.
* **AgentOps Monitoring Access Restrictions**: The AgentOps dashboard, conversation log export lists, and download information all require monitoring permission, preventing members with only other AgentOps permissions from reading organization monitoring data.
* **Preventing Forged Hook Verification Records**: Conversations no longer accept user-forged Hook execution results that could bypass personal data masking or input protection. Execution records are also no longer mixed into the assistant's fixed prompt content.
* **Stronger Identity Verification for LINE Web Pages**: When Web Chat is opened within LINE, visitor identity is accepted only after verification by LINE. LINE account information supplied by the web page itself is no longer trusted.
* **Stronger Activity Log Read Permissions**: Viewing activity log event details, export lists, and download links requires activity log permission, preventing unauthorized members from reading organization operation information.
* **Provider Response Storage Settings**: Responses requests sent directly to OpenAI or through Azure disable the provider-side response storage option, reducing additional storage of conversation replies through this interface.
* **Restrictions on Rejoining Removed Members**: After an administrator removes a member from an organization, SSO login and directory synchronization no longer add them back automatically. A new invitation is required to restore organization membership.
* **Stronger Information Masking in Assistant Teams**: When original user input is disabled for a node, attachment reading, parsing, and conversation log tools are also removed. Tool search no longer offers these restricted operations, preventing them from bypassing the masking scope.

### 🛠️ Bug Fixes <a href="#v2026-10-05-bug-fixes" id="v2026-10-05-bug-fixes"></a>

* **Knowledge Base Hybrid Retrieval Settings Fix**: When using a custom Elasticsearch knowledge base, connections supporting RRF are no longer incorrectly identified as unavailable, allowing hybrid retrieval to be enabled normally. A warning also appears if OpenSearch search configuration creation fails.
* **Deleting Connectors Together with Their Tools Fix**: When deleting a custom connector with the option to delete its tools selected, the deletion option is now sent correctly, preventing associated tools from remaining in the list.
* **Member Import Results Fix**: After batch imports of organization members or role members, results separately show counts for newly added and existing members, preventing existing members from being counted as successful additions.
* **Reply Streaming and Waiting Indicator Fix**: The first text segment of a reply appears as soon as it arrives, rather than briefly showing only a later segment. While a tool is processing and no text or cards are available yet, Web Chat continues to show a waiting indicator.
* **Duplicate Conversation Platform Sidebar Items Fix**: After opening a conversation from the support sidebar, the same conversation platform no longer appears multiple times due to pagination loading. Icons and channel labels also remain consistent.
* **Connector Authorization Status Fix**: Connectors using service accounts or requiring no authentication display an available status. They no longer ask members to authorize individually or show inapplicable deauthorization actions.
* **Admin Interface Acceptance Issues Fix**: 1. Meeting translation model searches now find matching names; 2. Usage statistics no longer accept future dates; 3. The delete button for the system administrator role is disabled in advance; 4. Activity logs no longer display inapplicable read events; 5. Access to the conversation export page follows monitoring permissions.
* **Tool Creation Dialog Fix**: When starting creation from a tool type card, the dialog no longer disappears immediately due to a page reload, allowing tool settings to be entered normally.
* **Gateway Trend Date Fix**: Trend charts in the AI Gateway overview display dates according to the selected time range. Queries spanning multiple days show both date and time so data from different dates does not appear to fall on the same day.
* **Knowledge Base Selection State Fix**: 1. Available knowledge bases update when search criteria change; 2. Rapid conversation switching retains the latest list, without older data returned later overwriting it; 3. MaiGPT can send messages directly when no knowledge base is selected and no filters are set.
* **Workbench Conversation Continuation Display Fix**: The Work workbench now displays context usage and continuation prompts for assistants other than MaiGPT and for admin test conversations. Creating a continuation conversation from a summary card keeps users at the original assistant entry point.
* **Disabled Skill Options Fix**: Skill menus for AI assistants, MaiGPT, and document quick actions list only enabled skills, preventing submission failures caused by selecting disabled organization skills.
* **Large Referenced File Preview Fix**: When opening large documents from knowledge base citations, file views, or monitoring, the preview wait limit increases from 20 seconds to 300 seconds, reducing timeouts before documents with hundreds of pages finish loading.
* **File Generation Failure Display Fix**: If an assistant is denied permission to start writing a file or writing fails, subsequent explanations remain visible as conversation text. They are not placed in a file card, and the file is not incorrectly shown as complete.
* **Connecting Existing WhatsApp Numbers Fix**: When connecting an existing WhatsApp Business App number to a conversation platform, the system correctly recognizes the coexistence connection completion event and no longer incorrectly shows registration as incomplete.
* **Role Section Preview Fix**: When only some sub-permissions are selected for a role, the section preview reflects the permissions actually available, matching the sections members can see after logging in.
* **File library version history entry fix**：The shared space root directory now includes an entry for version history. When switching workspaces or spaces, history stays separate for each scope, and older data that arrives later no longer replaces the current view.
* **Fallback display when maps are not configured**：When Google Maps connection details are not configured, the Web Chat map canvas displays an unavailable notice while keeping the location list and filters accessible, avoiding repeated loading failures.
* **Connector authorization status refresh fix**：Opening a conversation no longer triggers repeated connector status queries for the same identity. After authorization is completed, the latest status is fetched again so the connector no longer appears unauthorized.
* **Browser tool long operation fix**：When the assistant loads a slow webpage, it follows the configured wait limit, and long operations no longer lose the browser connection. Other conversations rejected because of usage limits no longer get stuck waiting for the browser to close.
* **Duplicate website assistant Widget fix**：Website assistants created through Ask AI reuse the assistant's built-in web Widget instead of creating a second identical entry. Settings customized by customers are preserved.
* **Delegated assistant interactive content fix**：When an assistant team delegates to a sub-assistant, the sub-assistant's question options, quick replies, rich cards, and files reach the user instead of being lost when results are handed back.
* **Video generation cost aggregation fix**：Conversation Credit consumption totals and conversation history exports now include video generation costs, giving administrators a complete view of consumption when checking usage.
* **Tool search authorization guidance fix**：When an assistant using tool search encounters an MCP tool that the contact has not authorized, it can still provide an authorization link, allowing the contact to complete the connection instead of only being told the tool is unavailable.
* **Keycloak connection error messages**：When SSO login fails because Keycloak certificate verification fails, a recognizable connection reason is displayed, helping distinguish verification configuration issues from general login failures.
* **Activity log recording fix for web operations**：Organization operations performed by administrators on the web are now correctly recorded in the activity log, fixing cases where an operation succeeded but its corresponding event was missing.
* **Existing conversation model synchronization fix**：When an administrator changes an assistant's model, existing conversations that inherit the assistant's settings automatically switch to the new model. Models explicitly selected by users within conversations are preserved.
* **Scheduled execution status fix**：Long-running scheduled tasks no longer start repeatedly. Tasks that exceed the execution limit are marked as failed, and running statuses left behind after unexpected interruptions are automatically cleared.
* **Provider credential preservation fix**：When editing a vLLM or OpenAI-compatible provider, changing only the name or other settings preserves existing credentials. Leaving the secret field blank no longer clears the API Key.
* **Automatic conversation evaluation connection fix**：When RAGAS automatic evaluation is enabled, consecutive scoring requests no longer log a connection error for each item because a shared connection has become invalid. Scores for each metric can be retrieved normally.
* **Tool failure reasons and trace fix**：When MCP or API tools cannot run because of insufficient permissions, connection failures, or other issues, the assistant and execution trace display understandable reasons. Failed calls are no longer marked as successful.
* **Model fallback failure handling fix**：AI Gateway fallback skips deleted self-hosted models. When no candidate model is available, it returns a clear failure reason instead of getting stuck on a model that no longer exists.
* **Blank response fix after team handoff**：After an assistant team hands off to the next node, the receiving assistant correctly identifies messages from earlier nodes. If its first attempt still produces a blank response, it tries once more.
* **Meta channel connection version update**：Messenger, Instagram, and WhatsApp connections and the Facebook login flow now use consistent, supported service versions, preventing channel failures caused by expired versions.
* **Presentation creation and image insertion fix**：Submitting through the presentation shortcut creates an editable presentation project. Generated images must actually be inserted into slides before completion is reported. Images not yet inserted are also available in the Insert Image drawer.
* **Model tool capability check fix**：Assistants using models that do not support tool calling no longer attempt to run tools. When deletion of a custom connector still in use is rejected, it is no longer incorrectly recorded as an authorization revocation.
* **Tool trace record link fix**：Tool nodes in execution traces now include the corresponding tool execution records, allowing administrators to review the inputs, outputs, and errors for each operation.
* **Assistant usage statistics fix**：Token usage from assistant responses is now correctly included in statistics. Hourly charts consistently group data by local time zone, preventing populated and empty time periods from being sorted incorrectly because of different time formats.
* **Knowledge base file move processing fix**：1. Files moved between knowledge bases are saved before processing starts, preventing moves from getting stuck; 2. evaluation report exports and test case generation follow the same order; 3. connector activation notifications are sent only after the application is saved, preventing missed notifications.
* **Disabled connector display fix**：Connectors disabled by the organization no longer appear among the connectors available in MaiGPT, including data returned by conversation updates and real-time messages.
* **Deleted fallback groups release credential references**：After a model fallback group is deleted, it no longer prevents the deletion of associated provider credentials. Unavailable settings are no longer counted as being in use.
* **Response emoji display fix**：When AI responses contain an index card emoji, it is displayed as ordinary text instead of being mistaken for a rich card and producing a loading failure message.
* **Bring-your-own-model usage recording fix**：When an organization disables Credits billing, assistant responses using models with its own credentials are still recorded in AI Gateway request logs, allowing administrators to review actual usage.
* **Traditional Chinese account message fix**：When using the Traditional Chinese interface, login failures and related account operation messages are displayed in Traditional Chinese instead of untranslated English prompts.
* **Duplicate monthly Credit allocation fix**：When plan changes and monthly Credit allocations run at the same time, the system rechecks the allocation time to prevent Credits from being added twice for the same month.
* **Claude conversation length calculation fix**：Available conversation length for Claude assistants is now calculated correctly according to model limits, preventing the first message in a new conversation from being incorrectly rejected as exceeding the length limit.
* **Video submission status confirmation**：The assistant must confirm the actual submission result from the video generation tool before reporting that a request has been submitted, preventing success claims before a generation job is created.
* **Bedrock response wait fix**：For conversations using Bedrock models with tool operations, waiting for the first response chunk and timeout handling are now consistent, preventing premature interruptions while the service is still preparing.
* **Azure non-GPT model response fix**：Non-GPT models such as Mistral on Azure now use compatible output length settings, fixing response failures caused by rejected settings after a conversation request was submitted.
* **Improved waiting after calendar synchronization failures**：After the first meeting calendar synchronization fails, the page reports the status and retries in the background. Reopening the page no longer gets stuck waiting for the same synchronization each time.
* **Office file output format fix**：When assistants generate Word, Excel, PowerPoint, or similar files, plain text is prevented from being written directly as a binary document, avoiding files that cannot be opened after downloading.
* **Conversation memory fix for human handoffs**：Visitor and human agent messages exchanged while customer service takes over are synchronized with the assistant. When control returns to AI, it can continue from the earlier conversation without overlooking information confirmed during the handoff.
* **Image Search Fix for Specified Filenames**: When users ask the assistant to find images by a specific filename or entity name, results are filtered by the original filename to avoid returning visually similar images unrelated to the specified subject.
* **LINE Authorization Link Preview Fix**: Previewing a contact's individual authorization link in LINE no longer consumes its one-time authorization eligibility prematurely. The authorization flow starts only after the user actually opens the link and confirms.
* **Parsing Fix for Special Presentation Charts**: When an uploaded PPTX contains unsupported 3D or other special charts, only those charts are replaced with placeholder content. The remaining slides can still be parsed, instead of the entire presentation failing.
* **Chinese PDF and Output List Fix**: When the assistant creates a Chinese PDF, it uses fonts that support Chinese characters and checks the finished file. The download list includes only the requested output files, without program files used during generation.
* **Traditional Chinese Preservation in Document Parsing**: In knowledge bases with Simplified-to-Traditional Chinese conversion enabled, text extracted from documents is converted to Taiwanese Traditional Chinese usage. This prevents Simplified Chinese content from appearing after text recognition in images embedded in PDFs. Code and links remain unchanged.
* **Thinking Setting Fix for GPT Reasoning Models**: When users connect directly to OpenAI GPT reasoning models and adjust reasoning effort, a compatible request method is used. This fixes the issue where specifying a reasoning level prevented the model from responding.


---

# 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/2026/10.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.
