📌 Release Overview

• 2026.40

• Unique hosted:September 30, 2026

Conduct GA

Features marked Experimental are off by default and enabled per customer on request.

What's New

• Admin New Codex model choices: Codex spaces can select GPT-6 Astra, GPT-6 Luna, and GPT-6 Sol.

• User Private Knowledge Base home: Conduct agents now receive the user's private Knowledge Base home as UNIQUE_HOME at /home-<userId>, with memory stored under user-memory. The process HOME remains the sandbox workspace.

• User Streaming tool-call arguments: Tool-call arguments now appear in the chat while they stream instead of waiting for the complete payload.

• Admin Operational diagnostics: The Debug side panel can show a bounded snapshot of the latest request sent to the Claude or Codex harness.

• Admin Experimental Persistent Pi sessions: Pi-powered spaces keep the Pi process running between messages by default, making follow-up messages faster. Configuration changes or runtime failures restart the process, and the setting has no effect when workspace persistence is disabled.

• Admin Experimental — Generated-file groundedness checks: With FEATURE_FLAG_GENERATED_FILE_HALLUCINATION_CHECK_UN_22029 enabled, every generated file receives a verdict: grounded, contradicted, ungrounded, inconclusive, or not checked. The check also reports how much of the file's supporting material it could inspect in the message Debug panel under Grounding check > generated_files; no mark indicating the grounding result is shown on the file itself yet.

• User Experimental — Generated-file sources: With FEATURE_FLAG_GENERATED_FILE_HALLUCINATION_CHECK_UN_22029 enabled, generated files are delivered as preview cards with their source chips shown beneath the card, including cards delivered before the final answer. Repeated links to the same file produce one download card per answer.

Improvements

• User Higher turn capacity: Spaces that still used the former 50-turn default now use 400 turns. Custom turn limits remain unchanged. An migration script has been applied to update the configurations.

• User Active-work time limits: Waiting for the user’s answer in an elicitation request no longer consumes the active-work limit. The maximum active work time of the sand box is by default 45 minutes. The maximum time to wait for an elicitation response is up to 30 minutes per turn.

• User Clearer timeout handling: Maximum-turn notices now appear after the work completes, and state both the elapsed time and the configured limit. Force-stopped turns close their open message segments instead of leaving the chat stuck in a working state.

• Admin Cache-aware session budgets: Max Session Tokens now counts cache-read and cache-creation tokens alongside input and output tokens. On models using prompt caching, a non-zero budget is reached sooner and may need to be increased.

• Admin Model-specific reasoning controls: Conduct settings now show only thinking levels supported by the selected Claude, Pi, or Codex model, with readable labels. Codex GPT-5.1 reasoning values xhigh and minimal are sent as high and low, the nearest levels supported by Azure.

• Admin Codex runtime metadata: Codex sessions now receive Azure-compatible metadata for configured GPT reasoning models, including context limits and supported reasoning efforts.

• Admin Citation-aware groundedness checks: The groundedness check now receives the answer with its citations intact, and each supplied source keeps the answer's footnote number. Set Hallucination Check Keeps Citations to off in the setup settings to restore the previous behavior.

• User Experimental Complete generated-file checks: Generated-file grounding now treats table rows and stat cards as claims, skips citation-only lines, and checks long files across multiple calls. A turn checks at most three files and makes at most three judge calls; files needing more than five calls remain unchecked, and files that cannot finish publish no sources until a later turn.

• Admin Cleaner Conduct settings: Improved descriptions of configuration parameters.

• Admin Knowledge Base materialization protection: Skill downloads now share a process-wide protection cap with concurrency 4, a 45-second deadline, and two retries. Operators can override these values with CONDUCT_KB_PROTECTION_CONCURRENCY, CONDUCT_KB_MATERIALIZE_DEADLINE_S, and CONDUCT_KB_RETRY_BUDGET; CONDUCT_MAX_CONCURRENT_KB_CALLS remains the legacy concurrency name. This improvement prevents out-of-memory issues caused by skills loading.

• Admin Workspace checkpoint lifecycle: Active-chat checkpoints are no longer removed prematurely by automatic file-retention cleanup and still expire after the full inactive retention period. Optional immediate removal of checkpoint mirrors on chat deletion or expiry is possible.

• User More reliable Pi conversations: Pi-powered Conduct spaces now run Pi 0.85.1 with improved multi-turn conversations, cancellation, and provider retry handling.

• Admin Configured context budgets: max_context_tokens now also caps when Pi compacts a conversation. Compaction uses auto_compact_pct of the smaller of the model window and max_context_tokens, so models with windows above 400,000 tokens compact earlier than before.

• Admin New Assessment opens the Conduct interview only when the space setting sowConfig.sowNarrativeConductEnabled is true and a Conduct space id is configured. The setting stays off unless an admin sets it.

  • Behavior change: Spaces that previously opened this interview because FEATURE_FLAG_ENABLE_SOW_NARRATIVE_CONDUCT_UN_24573 was on now open the standard assessment page until sowConfig.sowNarrativeConductEnabled is true.

Bug Fixes

• User Generated files render correctly: Generated files linked from an answer are now attached to the message and open as preview cards. A file linked more than once appears as a single download card, and an attached file no longer triggers an Output assembly failed banner merely because the answer mentions an output path that was not created.

• User Calendar and contact files arrive as downloads: Generated .ics, .ical, .ifb, and .vcf files are now delivered as working downloadable attachments instead of being silently dropped.

• Admin Large groundedness checks no longer fail: Long, heavily sourced answers no longer produce a red Hallucination Check Error because the judge prompt is too large.

• User GPT-5.1 Codex agents recover their skills: GPT-5.1 now receives a corrected compatibility profile instead of Codex's generic fallback, restoring coding-agent instructions and file editing while preserving supported capabilities.

• User Citations remain usable: Follow-up answers keep MCP citations as clickable references, web citations render instead of appearing as raw text, and web sources show a hostname and icon derived from the page URL when their stored title is missing. Icon lookups are limited to public hosts, so internal hostnames are not sent to third-party services.

• User Unsupported footnotes disappear: Footnote markers that have no registered source are removed before delivery, while genuine citations, repeated citations, and notation such as x squared remain unchanged. The Debug panel reports how many markers were removed, and the reference list now includes every source cited during the turn.

• User Connector questions reach the connector: Outlook and similar questions now use the connected tool instead of reporting that the service is unavailable. A failed shell command is no longer treated as evidence that the connector cannot be used.

• User Confirmed connector writes resume correctly: Connector writes that require confirmation now defer through elicit, wait for the user's response, and resume after confirmation instead of timing out.

• User Skills refresh: New or edited Knowledge Base skills take effect on the next message, and a skill changed during a conversation is re-read before the next answer. Newly added skills are available on the first discovery turn, and edits made within two minutes no longer return a cached earlier version.

• User Agent activity stays consistent: Answering a question during an agent run updates the existing Ask step instead of leaving duplicate or stale failed steps. Tool calls that finish after the assistant has narrated no longer create duplicate failed steps, and successful question-and-answer runs no longer appear in error.

• User Intent-only turns are handled: A turn that only says what the assistant intends to do without calling a tool is no longer silently marked answered.

• User Stall notices are consistent: A stalled turn now shows one warning instead of two contradictory warnings. A bug report is requested only when the turn produced no work.

• Admin Debug information is restored and more precise: Codex debug data again includes configured, effective, and last-turn approval settings. The Debug panel also reports unresolved citation markers by source type and separates turns that never searched from turns whose references failed to resolve.

• Admin Public API polling completes reliably: Conduct now records stoppedStreamingAt alongside completedAt, including for failure notices. Public API and SDK clients waiting with send_message_and_wait_for_completion no longer time out.

• User Context Memory updates need fewer permissions: Updating Context Memory from Conduct no longer requires Knowledge Base write permission.

Security

• Admin Retention covers Conduct data consistently: After the tenant retention window, Conduct now blanks tool-step logs, quality-assessment notes, citation details, elicitation prompts and responses, short-term memory records attached without a specific message, tool-call and tool-response payloads, and the Debug data keywords field. The module-selection identifier remains available after redaction.

• Admin Tool-call logs respect the shorter retention window: When TOOL_CALL_LOG_RETENTION_DAYS exceeds DATA_RETENTION_IN_DAYS, tool-call logs are pruned using the shorter window and a warning names both values.

• Admin Critical sandbox vulnerabilities are patched: Conduct sandbox images refresh operating-system packages to address published CRITICAL container vulnerabilities.

• Admin Sandbox temporary files are isolated: Pi temporary files, including Bash output spills and files written to $TMPDIR, are stored under .pi/tmp with private permissions and excluded from checkpoints instead of sharing the pod-wide temporary directory.

Chat Experience

🚀 What’s New (Business/End-User Focus)

✨ Highlights at a glance

• Scheduled tasks grow up Experimental — answers show which task produced them, and tasks can be edited, paused, or deleted right from the chat.

• Interactive HTML pages do more GA — users can upload files into the chat straight from a page, and admins can enable and sign in to connectors (Outlook, Salesforce, etc.) from dashboards in one click.

• Better document translation — PowerPoint and Word translations keep formatting, symbols, and every piece of text in place.

• New models — Grok 4.7, Gemini 3.8 Flash, GPT-6 Luna & Sol, and DeepSeek V4.1 Flash are available (subject to each company's allowed-model list).

• HTML previews now available for everyone GA — Users can now preview HTML contents generated in chat.

⏰ Scheduled tasks: manage everything from the chat Experimental

User What users can do:

• An answer produced by a scheduled task now shows that task's card beneath it — naming the task, its schedule, and the space it ran in — so it's always clear where a result came from. If the task was deleted since the run, the card says so.

• Tasks can be edited, paused, resumed, or deleted directly from that card, and clicking the card opens it for editing.

• Monthly schedules are more reliable: a task set for the 29th, 30th, or 31st now runs on the last day of shorter months instead of skipping them, and the preview shows the day it will actually run.

• Lots of polish: task search can be cleared with one click, the creation form shows helpful examples, tooltips are cleaner, long space names show in full on hover, and the model list matches what the chosen space actually offers.

• Spaces that only work on an uploaded document (such as a Document Translator space) are no longer offered when creating a task, since a scheduled run has no upload.

User Interactive HTML pages and dashboards

• File uploads from a page: interactive HTML pages can now accept dropped files and send them into the chat together with a prompt composed from the page's form fields. Files show a clear Processing → Ready status and behave exactly like files attached in the chat input.

• Pages can save data: an interactive page can update a linked Knowledge Base data file from a button — only when the page opts in and the user has write access to that file.

• Connector dashboards for admins: admins can list, enable, and sign in to MCP connectors (for example Outlook, Salesforce, Yahoo Finance) from HTML dashboards in chat — enabling and signing in happens in one click. Non-admins do not see the deactivated catalog.

• Connection status on dashboards is now accurate for every viewer (“Connected” / “Signed in” reflect the viewer's own state), and already-connected buttons are grayed out.

User 📄 Document translation quality

• PowerPoint translations keep punctuation, tabs, and line breaks in place, and no longer leave sentences untranslated because of unusual formatting.

• Symbol-only content, such as check marks in PowerPoint and Word tables, keeps its original glyphs while the surrounding text is translated.

• When the language model returns a faulty batch, the affected pieces are automatically retried in smaller chunks — so every text piece ends up translated and in its original place.

💻 Code Execution: simpler, more consistent results

• User Results from the Code Execution tool (files, images, interactive HTML) are now always shown as cards in chat and can always be expanded into the side panel — previously this depended on several separate toggles that have now been retired. Spaces that use Code Execution will see the expand button on every result card. (check out section below: Code execution results always show as cards)

• Admin The space setting that hides executed code now also hides the source code inside result cards — users then see only the result, with no way to view or download the code. The setting remains on by default, so nothing changes unless an admin turns it off.

• Note the breaking configuration change in the warning panel above: the tool itself is now enabled per company via codeExecutionEnabled. (check out section below: Code execution results always show as cards)

Admin 🤖 Model line-up

• New: Grok 4.7, Gemini 3.8 Flash, GPT-6 Luna and GPT-6 Sol, and DeepSeek V4.1 Flash are available. Companies that maintain an allowed-model list must add a new model to that list before it appears in the model picker.

• Retired: DeepSeek V4 Pro and V4 Pro 0813 (spaces move to DeepSeek V4.1 Flash — see warning panel above).

• Claude Opus 5.5 was added and then withdrawn again within this release — it is not available.

User 🛠️ Skills improvements

• Skill settings now show each skill's Knowledge Base folder location, so it's easy to find where a skill lives — the link is shown only to users who actually have access.

• Skills stored anywhere in a user's readable Knowledge Base folders are now listed, not just those in the dedicated skills folders. Skills in another user's personal folder stay hidden.

• Creating a skill now reliably starts a new chat in a space where skills are set up; the Create button explains itself when a user can't create skills.

User 🔍 Smaller preview features

• “Response stopped early” alert Experimental — when an answer halts due to a system interruption, the chat shows a clear alert with a “Try again” action. Pressing Stop yourself never triggers it.

• Document link graph Experimental — Markdown and HTML documents opened in the side panel can show an interactive map of the documents they link to; click a bubble to inspect and open a file, drag bubbles to rearrange.

User 🔒 Security and privacy GA

• HTML previews are now private: a preview generated from chat can only be opened by the user who generated it — preview links stop working for anyone else, and identifiers are no longer written to logs.

• For operators: HTML previews can optionally be restricted so rendered pages may only connect to an approved list of external hosts (off by default — all hosts allowed).

User ✅ Reliability and quality fixes

• Chat is more stable under load thanks to improved connection handling between internal services — fewer dropped requests during busy periods.

• Chats that finish in the background now show the unread indicator, and deleting a chat removes it from Chat History immediately.

• When an error stops an answer, the reply is closed properly — the Stop button no longer stays stuck on screen.

• Citations from PDF files uploaded to chat now open on the referenced page.

• Chat History bulk selection uses the standard selection toolbar and works correctly on mobile; several mobile layout issues (overlapping headers, unreachable menus) are fixed.

• Images pasted into the chat input are named sequentially (image1.png, image2.png, …); pasted files with a real name keep it.

• Files generated by an assistant no longer reappear as pending uploads after a page reload, and assistants can save user memories without errors.

• Cost information in Debug Info now includes the full model spend of a turn, even when a limit stopped it early.

• Numerous visual fixes: readable form fields on layered surfaces, consistent themed colors, aligned icons and tooltips.

🔧 Improvements

ADMIN Optional Internal Search / Web Search exclusivity: A company can now make Internal Search and Web Search mutually exclusive in its spaces, so internal data isn't sent to external web queries. When enabled, turning on Internal Search for a space automatically disables Web Search there.

User Code execution results always show as cards: Images and files created by the Code Execution tool always appear as result cards in chat. The feature flag FEATURE_FLAG_ENABLE_CODE_EXECUTION_FENCE_UN_17972 is retired and can no longer be set. Spaces that never turned it on now see the cards.

User HTML results open as an interactive card: HTML files created by the Code Execution tool always appear as an interactive card with a built-in code view. The feature flag FEATURE_FLAG_ENABLE_HTML_WITH_FENCE_UN_17927 is retired. Admins can still switch the card off per space with the enable_html_with_fence field in the Code display section of the space configuration.

User Open a code execution result in the side panel: Every result card (image, HTML, and file) now shows an expand button that opens the result in the chat side panel. The feature flag FEATURE_FLAG_ENABLE_CODE_EXECUTION_SIDE_PANEL_UN_18787 is retired. Spaces that run Code Execution but never turned the flag on now see the expand button on every card. The chat widget has no side panel, so it shows no expand button there.

Admin Code Execution is now a company setting: The Code Execution tool is controlled by the company setting codeExecutionEnabled instead of FEATURE_FLAG_ENABLE_CODE_EXECUTION_UN_17498, which no longer has any effect. Admins set it per company through the company configuration API. Operators seed it on backend-service-configuration with COMPANY_CONFIGURATION_CODE_EXECUTION_ENABLED="*:true" for a whole deployment, or "companyId:true;..." for single companies.

• • Behavior change: the setting is off by default. A deployment that runs Code Execution today must set the variable, or the tool disappears after the upgrade. See 2026.40 Infrastructure Changes.

Admin The "show executed code" setting now hides the code everywhere: The space setting "Enable display of executed code before the assistant message" (Code Execution, Code display) now also applies to the code execution result cards. With the setting off, the image, file, and HTML cards show only their result: no Content/Code switch, and no way to copy or download the source. The document and office preview side panels follow the same rule. The setting stays on by default, so nothing changes until an admin turns it off.

🐞 Bug Fixes

User Chat no longer freezes on large code execution results: A result card with a very large source, such as an HTML file with a big embedded script, could freeze or crash the whole chat tab. The card now builds the Code view only when you open it, and it skips colouring the source when the source is larger than 500 KB, showing a short "too large to display" message instead. The Copy action is hidden at the same limit. Download always stays, so you can still get the full source.

Administration / Governance

Added

• Admin Space Advanced Settings can auto-open the configured side-panel file in fullscreen. Enable it with Side Panel → Open in fullscreen. Requires document previews via FEATURE_FLAG_ENABLE_DOCUMENT_PREVIEWS_UN_20061.

• Admin Skills root folder selection now supports searching folders by name.

Changed

• Admin Space side panels can be configured by selecting a Knowledge Base file instead of entering a content ID.

• Admin Skills source headings and descriptions in Space Admin now match other space sections, with muted description text.

• Admin Personal skills and Additional sources in space Skills configuration are always shown when the Skills Sources card is visible. SKILLS_V2_ENABLED continues to gate only the Skills tab in user settings (with FEATURE_FLAG_ENABLE_SKILLS_UN_20042).

• Admin Space Skills settings show one Personal skills creation switch (enableSkillCreation). Existing personal skills always load.

  • Behavior change: enablePersonalSkills is removed from the Skills tool schema. Leftover saved values are ignored.

• Admin Skill tool sections in space settings use the same accordion as the rest of the space form.

• Admin Selecting a skills root folder now happens by clicking the folder. The checkbox is no longer shown.

• Admin Cost Management reports show cache tokens and their effective cost per 1M tokens alongside input and completion tokens, so the token breakdown accounts for all of Total Spent. Hovering a cache total shows its read and write split, and a new Space Type view (Unique AI Chat, Conduct, Agentic Table, …) breaks usage down per space type with a drilldown to assistants. Model Pricing lists the cache read and cache write rates. Requires FEATURE_FLAG_SAVE_MODEL_USAGE_DASHBOARD_UN_18889, off by default.

  • Behavior change: cached tokens on chat streaming requests were previously counted as regular input and billed at the full input rate. They are now billed at the lower cache rate, so reported spend and spend-limit consumption decrease for affected traffic.

• Admin In the skills root folder picker, clicking a folder row now expands that folder as well as selecting it.

Fixed

• Admin The back arrow on Create a new space returns to the spaces overview.

• Admin the spaces list in the admin side navigation no longer renders a space row's "..." menu inside the row itself, which produced invalid markup and browser console errors; the menu still appears on hover and can now also be reached with the keyboard.

• Admin Secondary badges use the secondary colour token. Additional source folder chips in space Skills settings also show a folder icon to match the design.

  • Behavior change: All BadgeV2 secondary badges pick up the new colour.

• User Users without admin access are redirected to Chat instead of a dead-end error page when opening the admin area; the admin home now opens the first section the user can access.

• Admin Space tool settings no longer show a Configuration section when it is empty or contains only hidden fields.

• Admin Space Skills Sources reopens when those fields have errors. Nested tool Configuration headings no longer collide with the space Configuration section.

• Admin Folders that contain files but no subfolders now say "No subfolders" in the skills folder picker, instead of claiming the folder is empty.

• Admin Improved the stability of the Chat Interactions and Detailed Chat Interactions analytics reports for large data sets. Reports covering many users and long date ranges previously could fail; they now complete reliably and generate faster.

• Admin Cost Management usage table shows the full model name without overlapping the next column, and lists each token column next to its cost column.

• Admin Selecting a folder in additional skill sources now also selects its subfolders.

MCP Hub Beta

Highlights: simpler tool approval (Always allow or Needs approval), safer defaults for new tools, and refresh that switches off tools whose definition changed.

What's New

• Admin Deferred tool confirmation for integrations: An integration calling an MCP tool that needs approval can set deferConfirmation to get an immediate answer instead of waiting for the user. The response carries confirmationPending and an elicitationId. Once the user accepts, the integration calls the tool again with that elicitationId within 5 minutes to run it.

Improvements

• Admin Simpler tool approval settings: Each MCP tool is now either Always allow or Needs approval, set on a single tool or on all tools at once.

  • Behavior change: Tools previously set to Ask-with-values become Needs approval. The feature flag FEATURE_FLAG_ENABLE_MCP_TOOL_POLICY_FOUR_STATE_UN_20686 is retired; the two-state setting is always on.

• Admin Safer defaults for new tools: A newly discovered tool starts switched off. MCP Hub tool cards flag every tool its provider has not explicitly marked read-only; that marking is advisory and controlled by the tool vendor. Tools already set to Always allow keep that setting until an admin changes it.

  • Behavior change: Once switched on, a tool that is not explicitly read-only asks the user for confirmation before running. Admins can still set it to Always allow explicitly.

• Admin Refresh Tool Settings for every company: Refresh Tool Settings in the admin Tools section is no longer behind a feature flag; FEATURE_FLAG_ENABLE_MCP_TOOL_REFRESH_UN_19255 is retired.

  • Behavior change: The Refresh Tool Settings button is now shown for every company with MCP, including companies that previously had the flag off.

• Admin Refresh switches off changed tools: When a connector is refreshed, any tool whose definition changed upstream is switched off, including when only its description changed, since descriptions are passed to the model. An admin has to review it and switch it back on before the agent can call it again. A tool that disappears and returns unchanged keeps its previous setting.

  • Behavior change: On the first refresh of an existing connector, tools that differ from their last known definition may be switched off for re-approval.

Bug Fixes

• User GitHub stays connected: GitHub MCP stays connected after its access token expires, so users no longer need to reconnect every day.

• Admin Refresh keeps your tool settings: Refreshing a connector's tools no longer discards the display names, tool usage instructions, or response format instructions an admin entered. A refresh that brings no changes from the connector leaves everything untouched instead of marking tools as updated. Display names also pick up the title a connector advertises.

• Admin Removed tools no longer block saving a connector: Connectors can be saved again after one of their tools is removed upstream. Previously the removed tool stayed enabled in the background and every save of the connector failed.

• Admin Unavailable tools on virtual servers: Virtual servers show tools their connector no longer offers with an Unavailable badge and a locked switch.

• Admin Tool changes reach chats right away: Tool changes from a refresh now reach spaces, chats and virtual servers without another tool having to be changed and saved first.

Security

• Admin Space tool settings are enforced: MCP tool calls made from a chat or message now respect the tools enabled for that space, including calls from Conduct and unique-cli.

Agentic Table

What's New

• User GA Bulk actions on Agentic Table sheets (the Bulk Actions button and the multi-select action bar) are now enabled by default wherever Agentic Tables are enabled.

• User Experimental Share the Answer Library from its Share button: Users with Can manage on the Answer Library can share it and grant Can edit or Can manage. The "Add to the Answer Library" option shown when a sheet completes is disabled for users without Can edit or Can manage on the library. Enable via FEATURE_FLAG_AGENTIC_TABLE_SHARE_LIBRARY_UN_23456.

• User Undo cell edits: Ctrl+Z (Cmd+Z on macOS) undoes the last manual cell edit in an Agentic Table or Answer Library sheet. Pressing it again steps back through the edits made since the sheet was opened.

• User Excel export now carries tags: The unformatted Agentic Table Excel export has a new Metadata Tags column right after the Question column, holding the sheet's tags and each row's tags as JSON. An exported sheet can therefore be imported into another space with its tags. The formatted export is unchanged.

  • Behavior change: in unformatted exports, every column after Question moves one place to the right.

• Admin Import tags from an exported sheet: Uploading an Agentic Table export (a file named export.xlsx) as a question file now also imports row and sheet tags from its optional Metadata Tags column; sheet tags in the file replace different values already on the sheet. Files without the column import as before, and spaces that already define a Metadata Tags column keep importing it as an ordinary column. The upload is rejected with the Excel row numbers, and nothing is imported, when a Metadata Tags cell is invalid, a row has tags but is otherwise empty, or sheet tags differ between rows or between files in one upload. If tags cannot be saved after the rows are imported, the activity says which rows lack them.

• Admin Public API to manage sheet access: New Public API endpoints list, grant, change and revoke user and group access on Agentic Table sheets. The list endpoint pages with skip=0 and take=50 by default (take max 1000).

Improvements

• User Bulk assignment of up to 50,000 rows: Bulk assignment now accepts selections of up to 50,000 rows.

• User Immediate feedback when a question is edited: Answer cells show the re-run loading skeleton as soon as a question edit starts a row re-run.

• User Share dialog shows each user's own access: The Agentic Table share dialog now shows each user's own access level instead of the highest level inherited from their groups. That level can be changed independently of group access, and group access is still shown on each group's own entry.

• Admin Faster Full Report Word export: The Full Report Word export no longer loads unused cell history, so exports of large tables complete faster. Integrations that need cell history pass includeLogHistory=true on the sheet or cell GET.

  • Behavior change: GET /public/chat/magic-table/:tableId/cell omits logEntries unless includeLogHistory=true. Sheet GET already required that flag.

• Admin Agentic Table Public API moves to version 2026-03-01: Public API /short-term-memory and /magic-table traffic, including POST /content/magic-table-sheets, now resolves to Unique API 2026-03-01 wherever unique-api forwarding is enabled. GET /magic-table/:sheetId keeps rowOrders cell filtering working for existing clients, including bracketed and large indexed rowOrders[] arrays.

• Admin Old sheet-access removal mutation retired: The GraphQL mutation deleteManyMagicTableSheetResourceAccess is no longer available.

  • Behavior change: integrations must use deleteManyMagicTableSheetAccessByResourceAccessIds to revoke grants by access entry id.

Bug Fixes

• User Row selection behaves as expected: Selecting rows and then doing something else, such as clicking a cell, changing a row status or opening a menu, now clears the selection instead of leaving it stuck behind the new action. Opening and closing a column filter keeps the selection, and applying a filter narrows it to the rows the filter kept.

• User Bulk metadata edits are reliable with collaborators: Bulk edit metadata stays open when another collaborator updates the same sheet; if selected rows are deleted meanwhile, they are removed from the selection and the user is asked to retry. Bulk metadata edits now appear immediately on every loaded row, not only the first hundred.

• User Stop really stops the run: Stopping an Agentic Table run now also stops the in-flight subagent (Conduct or chat), instead of leaving it generating after the sheet is idle. It also stops in-flight Knowledge Base and Library answers instead of waiting for those cells to finish.

• User Row tags are no longer overwritten: Row metadata now prefers a user-set value over sheet metadata, then over extraction. Sheet edits no longer wipe user or extraction tags, and extraction only fills keys a row is missing. Setting or renaming a key the row already holds replaces the existing value, and edits of inherited values now survive later sheet syncs.

• User Row numbers match the sheet: The row index now leaves the header unnumbered and starts the first question row at 1. It no longer briefly shows 2 for the first question while a new sheet is still processing.

• User Edit with AI polish: Edit with AI now appears only on answer cells, copies a proposal with the standard clipboard path, and uses the chat checkmark on Accept. Question and follow-up bubbles stay right-aligned (min 320px, max 75% of the dialog), and question bubbles use the space theme background. Enable via FEATURE_FLAG_AGENTIC_TABLE_EDIT_WITH_AI_UN_24386.

• User More reliable question extraction: Full-document question extraction now retries once with chunked extraction when it returns no questions or hits a recoverable request error, instead of ending the run or leaving the table empty. Excel workbooks processed sheet by sheet no longer skip a sheet that is too large or fails; the whole workbook is extracted with chunked extraction instead. Applies only to spaces that enable useFullDocument (with useHierarchicalExtraction) or useXlsxPreprocessing in the question extraction configuration; all three are off by default.

• User Answer Library name stays fixed: Answer Library sheets always display as Answer Library, and their name cannot be changed.

• User API responses match the previous version: On the 2026-03-01 API, POST /messages/:id/stop now returns the message's references instead of null, and magic table artifact responses include userAbortedAt on the nested sheet, both matching 2023-12-06. POST /content/magic-table-sheets now accepts the top-level context field sent by library ingest.

Recordings

Fixed

• User The recording bot now joins the call you invite it from, including Teams meeting-ID (/meet/) links, and retries a failed invite when a transient network error interrupts its first connection attempt.

• User Saving a bot name that Microsoft Teams will not accept as a display name, such as one containing parentheses, now shows an error instead of being accepted and leaving the recording bot unable to join meetings.

Language Models

ADMIN Claude Opus 5.5

Claude Opus 5.5 is now available through LiteLLM via Anthropic and VertexAI:

• litellm:anthropic-claude-opus-5-5

• litellm:vertex-claude-opus-5-5

ADMIN deepseek-v4-pro and deepseek-v4-pro-0813 - Retirement & Migration

Due to the upcoming retirement of deepseek-v4-pro and deepseek-v4-pro-0813 on October 29th 2026, all spaces and AI templates previously using one of these models have been migrated to litellm:deepseek-v4-1-flash.

ADMIN Gemini 3.8 Flash

Gemini 3.8 Flash is now available through LiteLLM via Google Cloud:

• litellm:gemini-3-8-flash

ADMIN GPT 6 Luna/Sol

GPT 6 Luna and Sol are now available through Azure (DataZone Standard deployment):

• AZURE_GPT_6_LUNA_2026_0922

• AZURE_GPT_6_SOL_2026_0922

ADMIN Grok 4.7

Grok 4.7 is now available through LiteLLM via xAI:

• litellm:grok-4-7

RAG

What’s New (Business/End-User Focus)

• User Experimental Link graph for Knowledge Base files: Markdown files opened in the Knowledge Base side panel can show a graph of the files they link to, the same view already available in chat. Enable via FEATURE_FLAG_ENABLE_DOCUMENT_LINK_GRAPH_UN_26282 for the company.

Improvements

• Admin Experimental Metadata extraction handles longer documents: AI metadata extraction in the Knowledge Base now accepts documents of up to 100,000 tokens. New folders default to 10,000 tokens, and existing folder settings stay as they are until an admin updates them. Set it in the folder's ingestion settings; requires FEATURE_FLAG_ENABLE_AGENTIC_METADATA_EXTRACTION_UN_15619.

• Admin Knowledge Base API moves to version 2026-03-01: Public API requests to /folder, /content, /chunk, /search and /ingestion-metrics are now answered by Unique API version 2026-03-01 wherever unique-api forwarding is configured.

  • Behavior change: these routes are no longer served by the legacy Public API.

Bug Fixes

• User CSV files with multi-line cells import fully: CSV files whose quoted fields span several lines now ingest as complete rows. Table row counts are no longer inflated.

• User PDF ingestion with Docling completes: PDFs processed with the Docling reader now finish ingesting.

• User Star ratings in PDFs are kept: Four- and five-star ratings that are read as asterisks during PDF ingestion now keep all their stars instead of being reduced to a single star.

  • Behavior change: documents ingested before this fix keep the old text until they are re-ingested.

• User Subfolders of hidden folders stay visible: Folders created inside a hidden Knowledge Base folder now stay visible, so their contents can be listed. A folder is hidden only when that is set explicitly.

  • Behavior change: subfolders of hidden Knowledge Base folders are no longer hidden automatically. Existing subfolders that inherited the hidden type are made visible again.

• User Deleted files leave nothing identifying behind: When the document-retention job removes a file, its name, description, source URL and other identifying metadata are now removed as well.

• User Nested metadata is readable: File details and the chunk view now show nested metadata as dotted fields.

• User Knowledge Base side panel polish: The move action in the side panel is now labeled Move and uses the current download icon. File previews show the full first page when FEATURE_FLAG_ENABLE_KB_FILE_PREVIEW_UN_23201 is enabled.

• User Clear error when folder creation is not allowed: Creating a folder through the Public API now returns 403 when the caller has no write access to the parent folder.

  • Behavior change: these requests previously returned 500 on API version 2023-12-06 and 422 on version 2026-03-01.

• Admin Experimental Revoking file management keeps admin rights: Knowledge Base administrators can revoke file management access on root-visible files without losing their own implicit management rights. Requires FEATURE_FLAG_ENABLE_CAN_MANAGE_ACCESS_UN_11608 and FEATURE_FLAG_ENABLE_KB_ROOT_VISIBILITY_UN_23074, both off by default.

• Admin Experimental Clearer message when a view grant is blocked: When a user cannot receive a Knowledge Base view grant, the message now names the knowledge read role. Applies only where FEATURE_FLAG_ENABLE_CAN_MANAGE_ACCESS_UN_11608 is enabled.

• Admin Experimental Space folder pickers show only granted folders: With root Knowledge Base visibility enabled via FEATURE_FLAG_ENABLE_KB_ROOT_VISIBILITY_UN_23074, the folder pickers for a space's internal search show only folders the data administrator was actually granted. The Knowledge Base management view is unchanged.

• Admin Large uploads no longer overload ingestion: Uploading a large file no longer holds the whole file in the ingestion service's memory, which could exhaust the service. Uploads above the maximum size are rejected with HTTP 413. The limit defaults to the existing 2 GiB platform ceiling and is set with MAX_UPLOAD_FILE_SIZE_BYTES.

  • Behavior change: uploads above the limit were previously accepted and buffered; they now fail with HTTP 413.

• Admin File retention cleanup removes everything: Retention cleanup now also deletes the PDF preview generated for Office documents, not just the original file. Expired chat files are removed from the database entirely once their retention window passes, instead of staying behind as an emptied-out record. Chat file removal requires an operator to set FILE_RETENTION_IN_DAYS, which is 0 (disabled) by default.

Security

• Admin Ingestion images patched: The ingestor image patches a published HIGH-severity Python dependency vulnerability. The ingestion-worker image patches published HIGH and MEDIUM .NET runtime vulnerabilities, and its email sanitizer patches two published HTML sanitization vulnerabilities.