📌 Release Overview

• 2026.34

• Unique hosted: August 19, 2026

Chat Platform

🔧 Improvements

User Jump to latest message: A button now appears when you're scrolled away from the bottom of the chat, letting you jump straight back to the latest message.

Admin Experimental Existing service users can now be migrated into Unique: Administrators can run an opt-in, company-scoped migration to sync service users that already exist in Zitadel into Unique's user system. The migration can be rerun safely and should be executed once per company when enabling FEATURE_FLAG_ENABLE_SERVICE_USER_API_KEY_UN_22696. See Service Users & API Key Scoping for instructions.

Admin GraphQL operation names in the network log: An opt-in env var can append the GraphQL operation name to the request URL, so calls are easier to find in the browser network tab. It’s wired to GRAPHQL_OPERATION_NAME_IN_URL_ENABLED through the Helm charts.

🐞 Bug Fixes

User File upload spinner: File badges in the chat input no longer keep spinning after the upload has finished.

User Chat auto-scroll: The chat again follows new messages while they are generated, including when the previous navigation layout is on.

User Internal Search reference badges: Internal Search reference numbers are easier to read because the badge color is stronger.

User PowerPoint preview: .pptx slides stay centered in the side panel viewer when the thumbnail sidebar is open.

Admin Connect to Knowledge Base checkboxes: The Select all checkbox in the Connect to Knowledge Base dialog now lines up with the file list.

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

User Experimental Shared Chats — share a conversation with colleagues: Chat owners can now share a chat directly from the chat header. Pick individual colleagues or groups in the share dialog and copy a unique link. Recipients open a read-only snapshot of the conversation, including answers and references, exactly as it existed when the link was created.

• Immutable snapshots: every copied link captures the chat at that moment. Messages added later are not visible through earlier links, and creating a new link never invalidates previously shared ones.

• Controlled audience: who can be picked is governed by the company sharing policy, either members of the chat's space (including individual members of groups that have space access, such as "All Company") or the entire company.

• Continue privately: recipients can fork a shared snapshot into a new private chat in their own history and keep working from there.

• Safe by design: recipients must be authenticated, uploaded files appear as locked, non-downloadable file names, and knowledge-base references only open for users who are independently authorized to access them.

• Revocable, with a bounded lifetime: owners can see everyone a chat is shared with and revoke access per recipient at any time. Links live only as long as the chat itself — once the chat is deleted or expires, its share links stop working.

This feature is disabled by default. Enabling it requires both the feature flag FEATURE_FLAG_ENABLE_SHARED_CHATS_UN_19869 and a company sharing audience set via the chatSharingPolicy company configuration (spaceMembers or allCompany). With only the flag enabled, sharing remains hidden for users and administrators alike. See the infrastructure changes subpage for details on how to setup the company setting for the chatSharingPolicy.

Admin Experimental Redesigned banner notifications: Administrators can schedule a company banner with start/end date and time, type, title, optional link, and a live preview. The banner appears as a full-width bar at the top of chat, knowledge, and admin. It's disabled by default and can be enabled via FEATURE_FLAG_NEW_NOTIFICATION_BANNER_UN_23900.

RAG

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

Admin Experimental Company-wide Knowledge Base visibility for data admins: Users with the chat.data.admin role can see all company Knowledge Base folders and files that already have at least one access grant, without needing per-folder access. Visibility is read-only: seeing a folder does not grant upload, move, rename, delete, or access management. General admin (chat.admin.all) is not a substitute. Chat search results, user-memory folders, and Skills trees stay access-filtered.

This feature is disabled by default. It can be enabled upon request via feature flag: FEATURE_FLAG_ENABLE_KB_ROOT_VISIBILITY_UN_23074.

Admin User name on analytics exports: User-focused analytics report exports now include a userName column, consistent with the Feedback export, so activity can be attributed to a person.

🔧 Improvements

User Faster search when scoping to a large file subset: RAG search no longer slows down severely when a space is scoped to a large subset of files (for example about 1,000 files in a much larger knowledge base). Responses stay usable instead of timing out.

User Faster folder tree for spaces with dynamic scope rules: The Select Files / folder tree loads faster for spaces whose scope rules are passed dynamically, instead of falling back to the slow unoptimized query path.

Admin Clearer ingestion failure statuses: Failed ingestions now report a more specific status when the cause is known (for example timeout on image-heavy documents), instead of a generic failure. This makes it easier to see why a file did not ingest from the status alone.

Admin PDF form fields and table headers ingest correctly: PDF ingestion no longer treats the first table row as a header when it is not one. Dropdown form fields can now include their full option list, not only the selected value. Expanding form options is off by default and can only be enabled via API.

This config is API-only; it is not available in the Knowledge Base UI. Enable it with pdfConfig.extractFormOptions: true in the Knowledge Base folder or RFP agent ingestion config, and in the space Advanced Settings.

Knowledge Base folder / RFP agent ingestion config:

{
  "ingestion_config": {
    "pdfConfig": {
      "extractFormOptions": true
    }
  }
}

Space Advanced Settings:

{
  "ingestionConfig": {
    "pdfConfig": {
      "pdfReadMode": "DOC_INTELLIGENCE_DEFAULT",
      "extractFormOptions": true
    }
  }
}

Agentic Table

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

User Can Answer role is now GA: Removes the Can Answer feature flag so the role is always available for share/assign flows.

🔧 Improvements

User Reduce filter checkbox flicker: Stops the table from fully re-filtering and flickering on every individual filter checkbox toggle so multi-select feels smoother.

User Idle debounce for new-sheet assignment emails: Adds an idle debounce before sending assignment emails for new sheets, so notifications wait until assignment activity settles. Defaults to 3 minutes and can be configured.

User Richer Answer Library ingest metadata: Stamps additional source and questionnaire metadata on library ingest and keeps it through re-verify for better traceability.

🐞 Bug Fixes

User Empty Agentic Table after questionnaire upload: Fixed a case where an Agentic Table sheet stayed empty (header only, no question extraction) after a questionnaire uploaded and ingested successfully, with no user-visible error that extraction never started.

User Answer Library header rows as data pollution: Source-sheet header rows are no longer pushed into the shared Answer Library sheet or KB scope as non-answer rows.

User Empty access list for users without display name: Access/share member lists no longer appear empty when assigned users have a null display name; a fallback label is shown instead.

Unique AI

🐞 Bug Fixes

User Prompt Language Preservation: UniqueAI now generates follow-up questions in the user's chosen language.

User Context Window Handling: UniqueAI now handles oversized prompts gracefully for Claude Sonnet 5 and Opus 4.7+ models, which use a different tokenizer compared to older models.

Language Model

ADMIN Gemini Flash 3.6

Gemini Flash 3.6 is now available through LiteLLM:

• litellm:litellm:gemini-3-6-flash: Directly through anthropic API

Conduct

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

User Created content is always delivered as a link: When Conduct creates or updates something — a file in the workspace, a knowledge base entry, or an item in a connected tool such as Jira or Confluence — the answer now always contains a clickable reference to it. If the agent produced a file but forgot to mention it, a reference is appended automatically, so generated results can no longer get lost.

Admin Version history for skills: Skill files in the Knowledge Base get a Version History panel showing prior versions, with download and one-click restore. Disabled by default; enable via FEATURE_FLAG_ENABLE_SKILL_VERSION_HISTORY_UN_23881.

Admin GPT 5.6 available for Conduct assistants in the model configuration.

Admin Opt-in native context compaction for Codex-based assistants: New assistant config (model_auto_compact_token_limit) lets the model summarize its own history before the hard context reset fires, preserving conversation continuity in long, tool-heavy sessions. No behavior change unless configured.

Admin Opt-in native context compaction for Codex-based assistants: New assistant config (model_auto_compact_token_limit) lets the model summarize its own history before the hard context reset fires, preserving conversation continuity in long, tool-heavy sessions. No behavior change unless configured.

🔧 Improvements

User Clear explanation when a run stops early: Runs that hit the turn limit, the cost budget, or an internal error now show a visible notice in the chat instead of appearing to go silent. A heads-up notice also appears when a run approaches its turn limit, and the agent can no longer declare success while its own plan is still unfinished.

User Better references from connected tools (MCP): Results without per-item detail (for example count-only queries) now get meaningful reference labels instead of a bare tool name, and answers citing several items from one search no longer produce dead reference chips.

User Fewer emojis: Conduct no longer decorates headings and answers with emojis by default; they remain available on explicit request/prompt.

Admin New Conduct spaces that use Claude now default to Opus 5: Opus 4.8 remains the fallback if Opus 5 is not available. Existing spaces keep the model they already have configured.

Admins who prefer up to 2× faster responses at slightly lower quality can select an earlier Opus version for new spaces.

Admin Context memory aligned with Space 2.0 policies: Conduct's context memory is now governed by the same rules as Space 2.0: gated by FEATURE_FLAG_ENABLE_CONTEXT_MEMORY_UN_22033, mutually exclusive with web search unless overridden. This avoids sending classified data to web-search without explicit configuration.

Behavior change: companies without the flag now have Conduct memory fully off.

Admin Correct model cost billing and attribution: Prompt-cache tokens are now billed at their proper discounted/premium rates (previously free or billed at the full input rate, most visibly on Anthropic models), and Conduct model spend is now attributed to the originating chat and assistant in cost reporting.

🐞 Bug Fixes

User Stuck "Thinking…" on failed or cancelled runs: Chats no longer hang on "Thinking…" forever when a run fails or is cancelled mid-turn; the turn now resolves with an explanatory message.

User Load more in long Conduct chats: "Load more" no longer skips older pages of the conversation.

User Memory updates on large knowledge bases: Memory updates ("remember that…") no longer hang the chat, and background memory loading no longer shows stray steps in the chat.

User False warnings removed: The hallucination marker is skipped when an answer has no references to check, and saving a skill no longer reports "Output assembly failed".

User CJK text in document previews: Chinese, Japanese, and Korean text renders correctly in generated Word/PowerPoint document previews instead of showing boxes.

User Shared chats hide process details: Shared chats no longer include Conduct's internal process details; recipients see only the conversation. Shared chats are rated Experimental (see top).

User Improved run resilience: Transient model-connection errors are retried automatically before the answer starts, and Codex-based assistants keep their session continuity after an error or cancellation.

User Reference chips and math formulas: Reference chips no longer break math formula rendering.

Admin Unsupported Conduct space settings hidden: Space settings that Conduct does not support (Sub-Agents, Custom Instructions) are now hidden for Conduct spaces.

Admin languageModel override on space message API: The languageModel override is now honored on the space message API endpoint, matching the existing endpoint's behavior.