Skip to content
Agent Fleet

v0.36.0

Agents’ skills can now be browsed read-only, the assistant moves past a backend held up by a usage or rate limit, and work items are marked when they have new activity.

New

Section titled “New”
  • See your agents’ skills, read-only: the folders each CLI reads, which copy wins when two share a name, and what each file says. Personal skills of Claude Code, Codex, OpenCode and Muse Code are under Settings › Skills, and a repository’s (Cursor’s too) under Skills… in its row’s menu. (#2792, #2795)
  • The assistant steps past a backend held up by a usage or rate limit. A chat turn that fails on a lasting limit before showing any output is re-run on the next usable backend in your priority list, with a notice naming both, and AI-assist one-off generations skip that backend too. A limit about to lift does not switch (the turn ends with its error), a compaction that hits a limit is not retried elsewhere, and memory review never moves to another vendor. (#2780)
  • Work items show new activity: a row updated since you last looked gets a dot and a bold title, and GitHub and Bitbucket rows show their comment count. (#2779)
  • Settings › Machine charts the workspace’s memory pressure beside the Agent’s response time. It only measures; nothing warns or stops a session yet. (#2516, #2837)

Improved

Section titled “Improved”
  • Find works in the PDF pane: Ctrl/Cmd+F opens the pane’s own find bar, which searches every page (a word wrapped across lines included) and moves between matches, and the highlight no longer hides the word. (#2822)
  • The work item Merge button is offered to GitHub App sign-ins too, when GitHub reports that you can write to the repository. Fine-grained tokens still do not get it. (#2753)
  • “Review again in ‹reviewer›” is offered on every device for reviews started from this version; a link made earlier still shows only in the browser that made it. (#2496)
  • A Claude conversation restored from the Agent’s backup says so in the transcript, since its latest part may be missing. (#2740)

Fixed

Section titled “Fixed”
  • A mistyped slash command left its chat bubble at Pending forever on claude sessions. Claude’s “Unknown command” warning now shows in the chat. (#2830)
  • Jira: Markdown bodies came out wrong and issue create -P dropped the parent. Comments and descriptions now keep code and bold next to Japanese text and a list after a paragraph, and the new issue gets its parent. (#2809, #2812)
  • opencode 2: the Agent instructions preview was empty, and the first turn of a session made with /new did not update the session’s Working / Ready status. (#2799, #2827)
  • Assistant chat: claude’s own auto-memory stayed on while Agent Fleet memory was on. (#2488)
  • A slash command sent with an attachment showed an internal instruction in its chip instead of the attachment. (#2823)
  • Smaller fixes: a Control Plane URL with an IPv6 address was refused by the egress allowlist, an MCP connection test that ran out of time was reported as the process exiting, and some Settings headings sat flush against the block above. (#2754, #2820, #2813)

Other changes

Section titled “Other changes”
  • Posting a comment on a work item is accepted only from the Console, by the member who owns the workspace, as merging already was, and each post is recorded in the audit log. (#2752)

Upgrade notes

Section titled “Upgrade notes”

Nothing to do.

Agent CLI pins: Claude Code 2.1.312, Codex 0.165.0, Copilot 1.0.105, Antigravity 1.4.8, Cursor 2026.10.22-27dcf5f, Kiro 2.35.1, OpenCode 2.2.0 and Muse Code 1.4.14-R7309.1.