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/2026-10-20.md.
One-click Runtime upgrade: a Runtime that is installed and Available on the Computer you are using now offers One-click upgrade, bringing the Runtime up to its latest version without updating tools one by one. The upgrade runs in the background through this Computer's local service, the button cannot be clicked again while it runs, Computer and Runtime information refreshes when it finishes, and both success and failure raise a notification. In the same batch, one-click install now works on Computers you own, not just the one you are using; a disabled or hidden entry explains why (the Runtime does not support it, you are not the owner, the Computer is offline, or its local service has not confirmed support), the install guide entry was renamed to Manual installation guide, and one-click install now raises clear start and completion notices (HF-14202, HF-13470, HF-14068, HF-14191).
Independent queue timeout for Agents: Agent settings gained a queue timeout — 24 hours by default, adjustable between 1 and 168 hours, independent of the per-Task run timeout. Queued tasks are no longer dropped by the old fixed 2-hour queue cap: they keep waiting within the queue timeout and start normally when their turn comes (HF-14213).
Choose where a re-assigned Agent runs: assigning a task to an Agent or Crew that already has chats on it first asks Run in the existing conversation, Run in a new conversation, or Cancel; choosing a new conversation makes it the executor's primary chat on that task, and the Assignee tag follows. With no prior chat there is no prompt — the assignment goes straight through (HF-14194).
Credential guidance when Test connection fails: with a Runtime installed but no model credentials configured, Test connection no longer ends in a bare timeout: the failure card points out the credential or sign-in problem and shows that Runtime's login command (for example opencode auth login). Runtimes with credentials configured are unchanged (HF-13598).
Cleaner run-duration display: a conversation in the Working state shows the current run's duration live (for example "Working · 44m 43s"), and Completed conversations always show how long they took (HF-14198, HF-14045).
No more skeleton-screen flashes when switching chats: switching between task conversations, or between standalone chats (including Assistant chats), no longer flashes a loading skeleton in the content pane (HF-14167).
Agent cards on hover in Agent pickers: hover an Agent's avatar in an Agent selection list to see its Agent card (HF-14163).
My Tasks reopens the last category: entering My Tasks again loads and renders the category you visited last, instead of loading All first and then jumping — faster and cheaper (HF-14199).
Inbox status changes stay on the first screen: clicking a status-change message in the Inbox opens the task with the status visible on the first screen, instead of scrolling down to the activity area (HF-14159).
Images and videos in activity open directly: clicking an image, video, or other media in task activity opens the media itself, instead of expanding the containing topic (HF-14156).
Chat dialogs show the Agent's default working directory: when a chat dialog's working directory follows the Agent and the Agent has a specified default directory, the dialog shows the directory name too (HF-14200).
More natural message timestamps: message times in Agent chats read as full forms like "15 minutes ago" instead of "15 minutes" (HF-14179).
Smoother workflow editing: opening a workflow detail now starts with the log panel collapsed (pressing Run expands it so results are visible), and the canvas auto-zooms to a readable size centered on the trigger, so many nodes no longer shrink into a clump (HF-14008).
Steadier idle-machine scheduling: host memory admission in the local service was reworked (tighter reservations, an idle floor, and tiered budgets), so small-memory or heavily used idle machines are no longer misjudged as having no capacity — ending tasks stuck at "Queued" — while Agents that fill memory no longer get OOM-killed (HF-14175).
Server performance and stability work: reduced execution counts and polling scans for high-frequency SQL around heartbeats, task claiming, session keep-alive, IM outbox, and the inbox; smoothed space and organization cascade-delete spikes and the run-trace index lifecycle, lowering database pressure in large Spaces. Follow-ups from the 10.19 production incident review shipped in this batch (HF-14176, HF-14119, HF-14152, HF-14233).
Test connection failing with "ping request not found": fixed some Runtimes failing Test connection outright with the error "ping request not found" (HF-14254).
Model selection and display mismatch: fixed the model picked in a standalone chat's composer not taking effect (the default model was still used), occasional errors when switching models in the GUI, and the model dropdown not showing the default selected model list for non-owners viewing an Agent (HF-14252, HF-14130, HF-14195).
Manage-storage cleanup lagged: cleanup by time or by task status in Manage storage waited for the next GC, and confirming a cleanup showed no in-progress hint; cleanup now starts immediately (HF-14207).
Custom-field settings dialog overflowed: fixed the settings dialog growing without a scrollbar when a field had many options, stretching the window vertically and pushing the title and footer buttons out of view (HF-14206).
Conversation state out of sync: fixed the sidebar conversation list showing Completed while the detail pane still said Working until a manual refresh (HF-14154).
macOS restart-update could not be retried: after Restart update failed, checking for updates and retrying still landed in the abnormal-update notice and only a quick client restart succeeded; retry now works (HF-14110).
Linear sync configs could not be deleted: deleting a sync config kept failing with 409 "task tracker link changed; refresh and retry", the binding survived refresh and retry (HF-14094).
Agent variant still showed "Inherit" for the model after overriding the Computer: after a variant overrode the Runtime under Computer & Runtime, the Model row still read Inherit instead of decoupling from the cloned Agent (HF-14037).
No rollback after a failed update handover: when the new daemon candidate was killed during an update handover, nothing rolled back and the daemon stayed stopped; rollback now happens (HF-14078).
Codex not re-discovered after fallback: after the Codex fallback path took effect, a running local service did not re-discover codex candidates immediately and Install Location showed a stale path (HF-14010).
Connector skills injected too broadly: CLI connector skills signed in on a Computer by other Agents were injected into every run of every Agent on that Computer, wasting tokens; they are now injected only into Agents bound to that connector (HF-14205).
