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/2026-11-03.md.
Invite members to a conversation: the conversation's ⋯ menu gains Invite members to this conversation — pick a member with an optional note; the invitee receives an invitation card in your direct message and joins once they accept. Transfer this conversation to someone else is renamed Transfer this conversation to another member and now completes only after the recipient accepts (HF-14839).
Read receipts in direct messages: messages you send in human DMs now show whether the peer has read them (HF-14794).
Per-computer network proxy: the Computer detail page gains network-proxy settings sharing one data source with the CLI and the desktop app's proxy configuration. Manual mode adds per-protocol HTTP / HTTPS switches and proxy authentication; follow-system mode auto-detects GNOME / KDE / /etc/environment on Linux, can prefer existing environment variables, and supports authentication layered on the system proxy (HF-14812).
Hide conversations: sidebar conversation rows support Hide — the row leaves the list until a new message arrives, affecting your list only. The row's ✓ icon becomes ⋯, and Hide replaces the old "Done" item in the context menu (HF-14834, HF-14844).
Notifications for direct messages: notification settings gain a Direct messages section controlling whether incoming human-DM messages notify; desktop notifications are on by default (HF-14795).
Avatar decorations: avatar settings let you overlay badges — the Agent's Runtime icon, the host OS icon, or the default model's brand icon — with hover tooltips explaining each; several can be shown at once (HF-14831).
One entry point for new task conversations: the task chat panel's "Start a new chat with an Agent" entry is replaced by Create new conversation, matching the ⋯ menu (HF-14806).
Unread accounting: image-only messages, invitation cards, and transfer invitations now count toward unread and last-message time like any other message (HF-14845); the unread dot on human-DM rows becomes a count badge (HF-14842).
Runtime switching in chats: the rules for switching the Runtime / Computer inside a conversation now follow the Agent's settings and its private / public visibility instead of a blanket block (HF-14804).
CLI install screen: leaner copy and a relocated "Regenerate token"; the install command offers Auto-start on boot / Manual start (boot by default); token expiry counts down by minutes then seconds; the daemon screen shows the startup method with a hint when it can't be changed (HF-14853).
Performance: the web app no longer re-fetches whole grouped endpoints in bursts (computers / provider-options 9× each) and caches large responses (HF-14886); fixed a new inbox-group slow-query fingerprint (HF-14849).
Interface polish: "Copy local working directory path" reads "Copy working directory path" (HF-14843); tool-research groups no longer auto-expand when a finished run is opened (HF-14852); single-line activity rows no longer show the Agent's job title (HF-14830); visibility pickers on the Agent settings page were refined (HF-14836, HF-14837); suggested Agent names avoid real or sensitive public figures (HF-14807).
Send-now queued replies produced no output: fixed a queued follow-up showing no streaming content after Send now moved it into a new conversation, until the conversation was reopened (HF-14800).
False "Agent offline": fixed sessions being told the Agent was offline while the run was actually healthy and processing (HF-14835).
Unclosable new-conversation dialog: fixed the dialog after creating a new conversation refusing to close until the page was refreshed (HF-14856).
Chat layout glitches: fixed images occasionally stacking over text and the composer shifting into the message history (HF-14802).
Default-model switching: fixed being unable to switch the default model inside a conversation after OMP's new default hadn't synced, with the menu flickering on every attempt (HF-14816).
Docs folder rename: fixed a Docs folder whose name contains a dot (like V1.1) being treated as a file extension and re-appended with ".1" (HF-14796).
Jira integration page permissions: fixed the organization integration page not disabling bind/import for non-admins and rewriting 403 permission errors into a generic message (HF-14787, Jira part).
GitHub MCP credentials: fixed credential creation failing with "The credential store is unavailable right now" (HF-14825).
CLI security: agent view no longer prints environment variables (including tokens / webhooks) in clear text (HF-14818).
Proxy egress 407: fixed saved proxy authentication and custom SOCKS5 not applying, so local Agent tasks stopped failing with 407 (HF-14904).
Workflow async Wait: fixed async Wait blocks not resuming on schedule, with runs marked failed while the pause record stayed (HF-14761).
Crew relays: fixed silent relay completions being misreported as failures and ACP result channels being lost (HF-14864).
Antigravity connection test: fixed the ping falling back to the daemon's working directory /, making the ACP side fail mkdir /.gemini on a read-only filesystem (HF-14871).
Task activity count: fixed every operation within one description update being counted in the activity number (HF-14792).
Editor tables: fixed deletion misbehaving when a description contains two tables (HF-14788).
External-account badges: fixed the Members → external accounts view and member-list badges not refreshing after an integration connection was removed (HF-14793).
Onboarding Runtime detection: fixed early Runtime detection not taking effect (still undetected after 2 minutes); "Starting" entries sort first (HF-14810).
Regression fixes: remote branches and Runtime lists no longer miss runtimeProviderKey case normalization (HF-14900); working-directory cleanup no longer issues a machine-wide deep delete when grouped runIds are empty (HF-14899).
CLI white-box fixes: task cancel-run --dry-run no longer cancels the run for real (HF-14771); skill create / files upsert no longer ignores --content-file, which left skill content empty (HF-14772); subprocesses no longer receive double-encoded credentials when proxy accounts contain special characters (HF-14824).
