For AI agents: the complete documentation index is available at /llms.txt, the full documentation bundle is available at /llms-full.txt, and this page is available as Markdown at /changelog/2026-10-26.md.
MiniMax Code runtime: the Runtime list now includes MiniMax Code CLI (mcode, via ACP), usable as an execution engine like other Agent CLIs; the sign-in guidance and error hints cover every supported language (HF-14358, HF-14418).
Conversation fork: chat sessions support forking a new session from any message (including AI replies) — the new session starts at that message as the branch point and inherits the full context up to it, the original session is untouched, and both can continue independently. Works for task conversations and standalone chats (HF-14246).
Authorization levels for Codex / Claude Code: conversations and Agents gain default authorization (permission approval) level settings, first for the Codex and Claude Code runtimes; in-run permission approvals and structured user questions share the same pipeline (HF-13867).
Agents can create Agents: the CLI / API restriction that blocked creating an Agent under an Agent identity is lifted — an Agent can now create other Agents mid-task; irreversible CLI commands also gained a double-confirmation hint in their help (HF-13845).
More languages: Sharkly adds Traditional Chinese, Spanish, Portuguese (Brazil / Portugal), Indonesian, and Vietnamese, and completes Japanese and Korean; language packs now load on demand (HF-14219).
"Create duplicate" entry in Agent settings: Agent settings add a "Create duplicate" entry next to clone, and the former "Copy" action is renamed "Clone" (HF-14511).
New-Agent flow recommends Git repositories: in the working-style dropdown, the Git repository option is badged "Best practice", with a tooltip explaining why — decoupling Agents from any single Computer so they can switch machines, parallel work via Git worktrees, and Sharkly caching pulled repositories so you don't spend time or tokens re-fetching (HF-14508).
Space Git repositories in Agent settings: the Agent's Git repository section adds a "Repositories from the Space" module — enable all or select a subset (none enabled by default); with all enabled, repositories added in Space settings later are enabled automatically, and enabled counts include them in the header (HF-11910).
Richer Agent list rows: each Agent row in the Agents list now shows the default model, Skills count, and connected tools count before the owner's name, each with its icon (HF-14373).
Search in "My Tasks": the board header adds a "Find in view" entry — click or press ⌘F / Ctrl+F to open and focus it, filtering instantly by title and task number, matching Space boards and views (HF-14517).
Crew conversations stay on task: communication during Crew task execution now lands in one conversation per trigger source — @ mentions inside a session reply in place in that task's conversation, @ mentions in comments reply in the comment's thread, and assignment-triggered runs reply in the task's own conversation, instead of spawning scattered new conversations; @-mentioned crew members also receive the full crew roster and leader identity (HF-14495).
Automation owner and conversation attribution: task-status-change and task-comment automations add a required Owner field (defaulting to "Triggered by"), and automation-triggered conversations are attributed to the owner; scheduled and webhook run details now render with the same conversation component as task comments; the initiator shown for automation-triggered conversations is the real trigger source instead of the task assignee (HF-14309, HF-14283).
Language switch no longer duplicates built-in data: after an organization switches its language, built-in task types, task statuses, and status groups are remapped to the new language instead of being duplicated into two sets — no more paired same-name statuses that need manual cleanup (HF-14503).
Model fallback and multi-provider switching: when model discovery fails mid-session or the model list is stale, the session automatically switches or falls back to an available default model instead of breaking (HF-14333); with multiple providers subscribed (opencode, opencode-go, …), same-named models can switch between providers and failures such as insufficient funds fall back automatically, no longer pinned to one restricted provider (HF-14465).
Agent errors reported to the server: when an Agent call fails, detailed error information (including the runtime and its version) is reported to the server for error statistics and product improvement, instead of only a one-line client hint (HF-13649).
Consistent page skeletons: pages missing loading placeholders (such as All Agents) gained skeleton screens, with unified styling, animation, and show/hide timing across pages (HF-14516).
Combined creator display: when a task has both a human creator and a creating Agent, the avatar shows a combined stack and the creator line reads "{Person} & {Agent}" (HF-14538).
Guidance after adding a custom field: creating a custom field in global configuration now opens a dialog — instead of a toast — where you pick which task types in which Spaces get the new field, so it's selectable in tasks right away (HF-14544).
Cancelled turns collapse by default: "Cancelled by user" turns in Agent conversations now collapse by default, folding intermediate narration and internal error steps inside; expand to inspect (HF-14547).
Enter sends the next queued message: when the input is empty and messages are queued, pressing Enter sends the queued message marked "↵", same as "Send now"; the queue panel icon and row were refined (HF-14553).
UI and copy details: the web download-desktop dialog distinguishes "macOS (Apple silicon)" / "macOS (Intel)" and fixes the Linux icon (HF-14522); profile copy fix (HF-14525); Lark moves to the end of the collaboration tools list (HF-14526); the runtime "may upgrade" hint reverts to its original copy (HF-14521); automatic runtime install failures now guide manual installation (HF-14545); the task conversation list drops the "Assignee" badge and adds an "Executing task" badge for assignment-triggered conversations (HF-14502); the permission-mode picker in task conversations moves to the right of the Runtime picker, matching standalone chats (HF-14507); refreshed conversation-status icons (HF-14244); the sidebar "More (n)" popover widened ~30% (HF-14490); "More runtimes" darkened so it no longer looks disabled (HF-14491); the Space skills dialog widened (HF-14514); clearer copy and tooltips for "Export ACP session", naming runtimes that don't support it (HF-14424); confirmation-dialog primary buttons gained a uniform loading state (spinner + aria-busy, 28 spots across the app) (HF-14461, HF-13911); the quick-reply tooltip now opens to the left (HF-14453).
Conversation state and lists: fixed entering a running conversation showing "stopped" first and only recovering seconds later (HF-14457); fixed a queued message cancelled before dispatch making the whole session look stopped (run steps vanishing), and conversations created on another phone or tab not appearing in the desktop sidebar until refresh (HF-14482); fixed sidebar conversation rows sometimes not showing the latest message (HF-14332); fixed switching conversations in the list occasionally not switching the right pane or showing a stale conversation (HF-14492).
Run timeline collapsing and scrolling: fixed the view jumping up when collapsing an expanded run while pinned to the bottom — the collapse header now stays in place (HF-14483); fixed collapsed runs leaving stale trailing whitespace when new output arrives instead of re-pinning to the bottom (HF-14488); fixed a later human mention returning first and clearing the previous Agent message's waiting placeholder (HF-14463); fixed stale queue reasons persisting in the task summary after the wait gate changes (HF-14481); fixed content flickering while scrolling up — image attachment rows re-mounting with oscillating heights and triggering a scroll-compensation loop (HF-14556).
No more false "thinking" when only mentioning a human: fixed crew task conversations showing an Agent "thinking" placeholder for ~20 seconds when a message @ mentioned only a human — no Agent was actually running (HF-14451).
Board fixes: fixed "My Tasks" crashing the whole view on "Group & sort" (useViewStore must be used within ViewStoreProvider, also in production) (HF-14567); fixed column headers keeping unfiltered counts after a custom-field filter (header "Done 2" with zero cards) (HF-14459).
Inbox fixes: fixed the mobile inbox cramming too many inline buttons into each row (HF-14549); fixed the "All spaces" filter option being impossible to unselect — the checkbox never changed and no list request fired (HF-13766).
Sign-in email codes: fixed the web sign-in page failing to send email verification codes with "Failed to send the code. Please make sure the server is running.", blocking email sign-in (HF-14477).
Duplicating an automation no longer opens a browser: fixed duplicating an automation automatically opening a browser tab (HF-14450).
Desktop fixes: fixed the macOS quit-confirmation dialog rendering at the wrong height on first open (HF-13525); fixed the status-bar menu not following the app's light/dark theme (HF-14422); fixed the Windows installer title-bar icon showing a red error-like badge — installer and uninstaller icons are unified (HF-14435); fixed the "Repair local service" dialog closing on any click under the All-settings route so the repair never completed (HF-14158); fixed one-click repair failing to recover an invalidated machine session (previously only a CLI reconnect could) — both recovery paths now converge (HF-14432).
Agent and skill fixes: fixed "Import from runtime" in Agent settings finding no local Codex / OpenCode skills (613 skills reported as 0) (HF-14420); fixed the "Import from this Computer" skill dialog showing "no available Computers" for about a minute before loading (HF-14475); fixed shared Agents returning 403 for image/screenshot chats by the shared-with user (text worked) (HF-14214).
Agent CLI permissions: fixed sharkly chat message list being blocked by a route-level 403 (EXECUTION_FORBIDDEN) during Agent execution — the endpoint now follows the user's permissions (HF-14504).
Offline Computers catch up automatically: fixed tasks assigned to an Agent on a CLI Computer while it was offline not auto-summoning when the machine came back online (HF-14473).
Integrations and imports: fixed Shortcut imports turning external assignees into a nameless, email-less "External user" that never backfilled after the invite was accepted (HF-14402); fixed the Linear import letting you proceed without selecting a target Space — now a required check (HF-14310); fixed Bugzilla imports putting the original description into comments and leaving the task description empty (HF-14542).
Display detail fixes: fixed the dropdown checkbox overlapping its label (HF-14519); fixed Docs rows missing icons and hugging the left edge, misaligned with other rows (HF-14449); fixed the environment-variable dialog's KEY focus falling back to row 1 when opening a specific row (HF-13570).
Server performance and stability: optimized the board filtered_tasks slow query (90 hits / 141.9s) (HF-14408); optimized the task transaction advisory-lock wait slow query (279 hits / max 29.4s) (HF-14444); optimized the crews table UPDATE slow query (42 hits / max 14.9s) (HF-14446); optimized the task/comment fuzzy-search query scanning millions of rows (40 hits) (HF-14447); fixed the GitLab project-list endpoint returning 502 after 20s timeouts (HF-14448).
