Release 2026.30

9 min read

Release Overview

  • 2026.30

  • Unique hosted:


Chat Platform

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

Admin Experimental 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.

This feature is disabled by default. It can be enabled upon request via feature flag: FEATURE_FLAG_ENFORCE_INTERNAL_SEARCH_WEB_SEARCH_EXCLUSIVITY_UN_22973.

🔧 Improvements

User Chat Interaction Detailed analytics report: The report now includes two new columns, “Tool names” and “Skill names” used in chat.

🐞 Bug Fixes

User Analytics report overfetching: A bug causing analytics reports to extend by one day beyond the specified date-range has been fixed, ensuring reports now match the exact date-range..

RAG

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

User Experimental Recency ranking to RAG search : Introduced native recency-aware ranking in both Qdrant and Elasticsearch search backends so that newer documents are scored higher during retrieval. The decay is applied directly at the database scoring level using the existing metadata.validAsOf date field, not as a post-fetch rescoring step, ensuring the most relevant recent chunks are surfaced even at large corpus scale. Decay parameters (half-life, weight, floor) are tunable, and the feature is behind an environment configuration.

Admin Recursive folder access removal skips protected subfolders: Removing Manage access from a folder and its subfolders no longer fails when a nested folder would lose its last Manage permission. Those subfolders are skipped and left unchanged, while the parent and other eligible subfolders are updated as requested.

Admin Restricted administrative maintenance operations to admin users only: Fixed a broken access control vulnerability where the createMaintenanceJob GraphQL mutation was accessible to any user with knowledge-base manager permissions, potentially allowing them to trigger destructive operations such as embedding model switches, collection resharding, and full index rebuilds across the entire tenant RAG knowledgebase. The mutation is now properly gated behind admin-only permissions.

🔧 Improvements

🐞 Bug Fixes

User Fixed Excel ingestion crash on workbooks with orphaned drawing relationships: Resolved an issue where .xlsx files edited by external tools that deleted worksheets without cleaning up their drawing references would crash during ingestion with a Cannot read properties of undefined error. The Excel ingestion path now gracefully handles missing or inconsistent drawing anchors, allowing text and table content to be extracted normally.

User Fixed chat upload failure for Excel workbooks with large content cells: Resolved an issue where certain valid Excel workbooks failed during chat upload because long narrative content cells were misclassified as table headers, causing the synthesized header to exceed the chunk token limit and trigger a generic parsing error. The ingestion logic now detects oversized headers and gracefully splits table content into smaller chunks.

User Improved folder name handling: Added a limit to folder name length(1000) to prevent issues when updating the platform. If someone tries to create a folder with a name that's too long, they'll be asked to choose a shorter name. In cases where folder names come from external systems, they're automatically shortened when needed to ensure everything continues to work smoothly.

User Fixed missing Confluence dates during ingestion: Resolved an issue where Confluence /date macros were dropped from ingested content. These macros are returned as empty <time datetime="…"> HTML tags, which were stripped during HTML-to-markdown conversion because they had no text content. Dates from the datetime attribute are now preserved and appear correctly in document chunks.

Agentic Table

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

User Experimental Notifications for Agentic Table: The users have now the possibility to send out notifications to the assignees inside the Agentic Table to inform them that they have been assigned to specific rows.

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

FEATURE_FLAG_AGENTIC_TABLE_ASSIGNEE_EMAIL_UN_20992 to send notifications

FEATURE_FLAG_NOTIFICATIONS_BELL_UN_21769 to see the notifications in app

To configure the email provider to send emails refer to Notifications

🔧 Improvements

Admin Configurable Section column format: The Agentic Table Section column is no longer fixed to {id} - {title}. Its format is now configurable per space via a Jinja template (sectionFormat in the question-extraction config), which can reference the section id and title, the question id/path, and can mix separators (e.g. 1.10 - Firm: General Information). The default output is unchanged. Takes effect on the hierarchical-extraction and XLSX-preprocessing paths.

Admin Backfill legacy answer-library file provenance: New admin mutation restores readable file names and the reference sidebar for answer-library files created before the June 2026 naming change.

User Clearer Agentic Table error banners: Reworded ingestion-scope, invalid-configuration, and sheet-busy error messages to be concise and table-specific for GA.

User Copy action generally available: The Agentic Table copy button is now enabled for everyone; its feature flag was removed.

🐞 Bug Fixes

Admin Hallucination Score column now respects hallucination.enabled: Fixed an issue where disabling table-level hallucination evaluation (hallucination.enabled: false) did not fully switch it off and the Hallucination Score column was still processed. The flag now acts as a master switch for all table-owned hallucination work, and the column is left empty when disabled.

User Question files no longer retrieved or cited in subagent answers: Fixed an issue where a subagent answer column could search and cite the sheet's question file (the file questions were extracted from) instead of only the intended source files, including on sheet rerun, when sources were added later, and on single-row rerun. Question files are now excluded from answer-cell search by default; a space can opt out by setting excludeQuestionFilesFromSearch: false.

User Sheet title length: Fixed missing validation that allowed My Sheets titles longer than 70 characters to be saved.

User Re-run button after export import: Fixed the Re-run button not appearing after importing a question file whose name contains "_export".

User Row-delete confirmation: Fixed the missing "Row deleted successfully" toast when deleting an Agentic Table row.

Unique AI

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

Admin New Models Added to Unique AI Approved List: GPT-5.6 Sol has been added to the approved model list. This model is now appearing in the Unique AI dropdown, provided it is enabled for your organization.

🐞 Bug Fixes

User Prevent MCP Response Size Failures - Fixed an issue where oversized MCP server responses could exceed the model’s context window and cause requests to fail. Large responses are now truncated safely with a clear notice, allowing interactions to continue reliably.

User Large File Support for Code Execution - Code Execution now supports Excel and CSV files larger than 8 MB, ensuring they reach the analysis container successfully. This prevents uploaded files from remaining inaccessible to the model. The new file size limit is set at 50 MB.

Language Model

Admin GPT 5.6 Sol, Terra, Luna

GPT 5.6 models are now available through Azure (only as global and data-zone-standard deployment):

  • AZURE_GPT_56_LUNA_2026_0709

  • AZURE_GPT_56_SOL_2026_0709

  • AZURE_GPT_56_TERRA_2026_0709

Admin Model retirement

The following models are deprecated and not further available after July 10, 2026:

  • litellm:qwen-3-235B-A22B

  • litellm:qwen-3-235B-A22B-thinking

  • litellm:glm-5.1

  • litellm:deepseek-r1

  • litellm:deepseek-v3-1

Admin Claude Opus 4.1 - Upcoming Retirement & Migration

Due to the upcoming retirement of Claude Opus 4.1 on August 5th 2026, all spaces and AI templates previously using litellm:anthropic-claude-opus-4-1 have been migrated to Claude Opus 4.8 (litellm:anthropic-claude-opus-4-8).

Admin Litellm’s keys pruning

The unused keys for the teams/companies now gets pruned every night

Conduct

Conduct

Conduct inside this release is still rated as pre-experimental and only recommended for use with the Claude harness and models.

Conduct is currently only available via Unique SaaS (MT, ST), Self-Hosters must first procure their infrastructure as outlined in Self-Hosting Conduct.

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

User Experimental MCP source citations with links and titles: Connector references in answers can open the cited item directly and show a readable title instead of opaque source IDs when links are available.

User Experimental Live thinking in the activity timeline: Reasoning now streams into the step timeline while the agent works, alongside clearer logs for commands, searches, skills, and MCP tool calls.

User Experimental Built-in document skills: Agents can create and edit Word, PDF, Excel, and branded template documents from the sandbox, with vetted bundled skills for common office workflows.

User Experimental Interactive HTML deliverables: Upgraded skill for dashboards, one-pagers, and interactive reports shown in the sidebar preview, with improved branding and anti-flicker behavior.

User Experimental Report an Issue in Conduct chats: A composer action copies privacy-safe debug data (prompts and answers redacted) so users can share diagnostics with support.

User Experimental Edit memory.md from chat: Users can edit their persistent agent memory file directly in the Conduct chat experience.

User Experimental MCP connectors on GPT/Codex/Pi: Email, Atlassian, Slack, Teams, and other platform connectors work natively for non-Claude Conduct strategies, not only Claude Code.

🔧 Improvements

Admin Experimental Claude Opus 4.8 by default: New Conduct spaces default to Claude Opus 4.8 with Opus 4.7 as fallback.

User Experimental Faster cold starts and reduced time to first token: Boot and turn latency improvements include Claude client pooling, CLI warmup, parallel workspace setup, and cached skills manifests.

User Experimental Faster output handoff: Unchanged deliverables skip redundant re-upload; users see “Assembling outputs” / “Outputs ready” on the answer while files are prepared.

Admin Experimental Longer autonomous runs: Maximum agent turns can be configured up to 1000 (default raised to 400) for multi-step document and analysis tasks.

User Experimental Clearer step labels with rich icons: Activity steps use present-tense copy (“Running a script”) and distinct icons while in progress and switch to past tense only when complete.

Admin Experimental Skip ingestion by default for new Conduct spaces: New Conduct spaces now default to not indexing uploaded chat files into the platform’s retrieval/search pipeline.

🐞 Bug Fixes

User Experimental Stuck or silent agent turns: A stream stall watchdog and per-chat turn serialization prevent hung turns and concurrent messages corrupting the same sandbox session.

User Experimental Missing references on multi-step answers: The UI now binds citations, references, and hallucination checks to the final answer segment when a turn has multiple answer blocks.

User Experimental Preface vs final answer rendering: Interim narration appears as preface segments; the terminal segment is labeled as the answer without duplicate bubbles.

User Experimental MCP calls hanging on Codex/GPT: Connector tool calls no longer stall for 60–120s; egress proxy and SDK config are forwarded correctly to sandbox MCP proxies.

User Experimental False “high hallucination” on MCP answers: Citation markers are stripped before the grounding check so verified connector answers are not flagged incorrectly.

User Experimental Input blocked after streaming: The chat composer re-enables as soon as the answer finishes streaming, instead of waiting for slow post-processing.

User Experimental Max turns with no explanation: Users get a clear warning when the agent hits its turn limit instead of the response stopping silently.

User Experimental Select Files in Conduct: The Select Files control is hidden in Conduct chats where it had no effect on agent search scope.

User Experimental Internal search when disabled: Agents are instructed not to run knowledge-base search when Internal Search is turned off for the space.

User Experimental User memory scope lookup: Sandbox gateway routing fixes allow user-memory root group resolution so memory features work reliably.

Admin Experimental Vertex AI Claude Code: Claude Code strategy works correctly when models are routed through Vertex AI.

User Experimental Unreferenced file chips removed: Output files the agent did not cite inline no longer appear as forced chips at the end of every reply.

User Experimental Elicitation and segment history: Multi-segment turn history and elicitation segments are preserved correctly for follow-up turns.

Infrastructure Changes

2026.30 adds durable sandbox checkpoint storage (via Azure Files) to drastically increase sandbox and turn boot time.

Operational hardening includes stale sandbox-claim eviction when the deployed agent image changes and lower sandbox CPU requests with alerting for cost optimization at no performance loss.

Last updated