2.3.5

What's new in 2.3.5

Work in more windows, steer split panes from the keyboard, and keep scheduled and background work under control. Zana 2.3.5 opens Salesforce Studio in its own window, adds Scheduler recovery and Squad worktrees, lets you choose which skills agents load, and makes background processes visible and stoppable.

New

• Scheduler recovery. Choose whether missed runs catch up once, catch up each missed time up to a limit, or skip. Enable automatic retries for temporary failures per schedule. New schedules start with recovery enabled; existing schedules keep their timing until you opt in.
• Scheduler Problems view. See missed, skipped and failed runs with reasons and recovery hints, then retry or dismiss them individually or in bulk. Agents and zcc schedule runs, retry and dismiss can manage the same run history.
• Squad worktrees and recovery. Run a Squad in a managed Git worktree, with defaults at the global, project and Squad levels. Retained worktrees offer Inbox actions to keep or delete them, or instruct a follow-up agent. Squad details let you relaunch from a retained plan, resume an existing Squad or recover a timed-out run when available.
• Correct Squad plans before continuing. An invalid provided plan pauses for an Inbox correction and approval flow, then resumes the same Squad. Scheduler and Goal runs save the approved correction back to their prompt or objective, with checks for concurrent edits. Squad Goals now offer inferred or provided plans.
• Skills controls for local threads. Choose On, Manual or Off per skill, filter by harness and state, and search by source or plugin. Skills remain On by default.
• Salesforce Studio in its own window. Open Agentforce Studio as a separate window with an activity bar, org context kept in sync across windows, and a title showing the agent, org and a Production or Unknown org badge. Agent cards and Preview trace steps open there directly, and Bring back returns the tab to the main window.
• Release and Tests lanes for Agentforce. Save, compile and publish agents, then activate or deactivate org versions, each behind a signed approval sheet. View any version read-only with a diff against your draft. The Tests pane runs YAML specs, pins a preview as a test, and offers a fix loop.
• Agentforce Studio workbench. An assistant rail, inline agent edits with per-hunk accept or reject, gutter comments, traces, scenario suites and chat cards. A collapsible explorer adds a Lightning Types viewer, and anonymous Apex runs directly after review.
• Split-pane shortcuts. Move focus between panes with ⌘⌥ and the arrow keys, cycle with ⌘⌥[ and ⌘⌥], jump with ⌘1–8, maximize with ⌘⇧M and close with ⌘⌥W. All are remappable in Settings → Shortcuts. Dividers snap to an even split, panes of archived threads close themselves, and "Open in split below" joins the thread menu.
• Edit queued prompts. Change a queued message before it sends. Later messages keep their order, and a conflicting change keeps your draft.
• Notification levels. Choose all, input-only or muted per thread, with a default for child threads. Phone notification previews strip markdown and redact secrets.
• Stop background tasks from their card. Each background task gets a Stop control, plus Stop all, even while the thread is idle. zcc thread background stop uses the same path.
• Storage & retention. An optional built-in plugin clears old thread files on each machine. It is installed disabled, every policy defaults to Never, and it never deletes threads.
• GitHub login on other machines. The GitHub plugin can forward your local gh login to non-primary machines, scoped to github.com. A machine's own settings win.
• Publish design docs to GitHub Pages. A Publish dialog exports a doc as a static site, pushes it and reports the Pages URL. Agents opened from a doc stay in its side panel.

Improved

• Organize design docs. Pin documents, find those with open comments in Needs me, and group the list by project or use a Recent timeline. Alternative proposals nest beneath their original document, and collapsed sections remember their state.
• Shared question controls. Pi input requests and plugin questions use the same conversation UI, with choice descriptions, multi-select, navigation, cancellation and retained drafts.
• Control Scheduler concurrency. Set the maximum simultaneous runs and choose automatic, queued or skipped overflow. Recovery catch-ups and retries queue when the limit is reached.
• Clearer Squad launch failures. Required skills and MCP servers are checked before launch. Missing or disabled requirements show their cause in the launch composer, Scheduler run or Goal iteration.
• Project identity in the titlebar. In a single-project view the titlebar name opens the project actions menu, and the rail shows a flat Sessions list.
• Plugin side-panel agents know their plugin. Side-panel conversations start grounded on the plugin and page they were opened from, stay out of the Agents list, and can be promoted into it. A fresh panel composer offers your last five conversations.
• Running-process warnings. Thread rows show how many processes are still running, and archiving or closing warns that it will stop them.
• Faster AI Harness settings. Opening the page reuses the startup harness check instead of re-probing every CLI; Check status still re-probes on demand.
• Steadier terminals and inbox. Product resets no longer blank and replay terminals, terminals keep their scroll position, and Inbox revisits skip skeleton flashes.
• Calmer plan cards. A closed plan card stays closed across new messages and reloads, and finished plans leave the timeline.
• Longer agent questions. Ask User Question waits up to seven days, and long tool calls keep their connection alive.

Fixed

• Voice input no longer fails before transcription because its timeout exceeds the host command limit.
• Windows titlebars provide working native minimize, maximize and close controls, with consistent height and spacing.
• Claude Code CLI Agents no longer have every tool call denied when a native policy hook has no decision; the CLI's own permissions still apply.
• Cursor's Auto model selects correctly, and ACP connection failures retain their original error details.
• Harness setting changes refresh providers from the saved configuration. Pi default models, stale OpenCode model selections and project-specific agent discovery recover consistently.
• Pi input, selection, confirmation and editor requests appear in the conversation and return answers or cancellation to the provider.
• Design Docs' Copy chat reference and Copy id actions work in the desktop app, and picker search fields receive focus when opened.
• Google Analytics sends compatible installation IDs and engagement events so collection is accepted correctly.
• Pi trusts the operating system's certificate store, so providers behind a corporate root certificate connect when Zana is launched from the Dock. A "Trust system certificates" setting controls it.
• Native skills are discovered on the machine that runs the agent, not only the primary machine.
• Threads created right after launch wait for their provider plugin instead of failing with "unknown thread provider".
• Shell reads such as cd dir; head file resolve to the right file, and Preview is hidden for files it cannot open.
• Agentforce Preview runs end cleanly when you switch files or close the panel, and Live sessions are released with their thread.
• Codex accepts none reasoning and no longer duplicates sub-agent follow-ups; ACP agents re-apply changed model, effort and tier on follow-ups.
• Long-lived plugin runtimes no longer hit "Plugin signal limit exceeded".
• Shared previews list the primary machine once, and pairing relay requests stay pinned to the product origin.
• Agents create design docs only when you ask for one.