Skip to content

v0.25.0

Latest

Choose a tag to compare

@faridun-m faridun-m released this 28 Oct 06:18
· 84 commits to main since this release
Immutable release. Only release title and notes can be modified.
29bb474

[0.25.0] - 2026-10-27

See how team context shaped each session in your status line and the new ox brief, meet Oxy in Codex and the Claude desktop app, and get sessions from worktrees, resumed conversations and commits into your Ledger again.

New

  • See how team context shaped your session, in the status line — Claude Code's status line now shows ● ox 3 discussions · 1 plan · 2 used: the team context your AI coworker was given, how much of it changed the work, and ● rec while the session records. In the Claude desktop app the same line appears above the prompt. ox doctor adds it without replacing a status line you already have; ox config set statusline off turns it off.
  • ox brief shows what SageOx gave you this session — the team context ox put in front of your AI coworker, what it actually used and how (read, cited, named in a commit), and whether SageOx would be credited as co-author. It updates while the session records; --me rolls up your recent sessions and --team shows team-wide totals only, never who read what.
  • A live brief pane beside the chat — /ox-brief opens this session's brief and every ox call in a pane in Claude Code, refreshing as calls happen; in the Claude desktop app it docks beside the chat on its own. ox init and ox doctor install it, and ox config set brief_pane off removes it.
  • Oxy in Codex and the Claude desktop app — ox pet enable now adds Oxy as a Codex terminal pet (in a terminal that can show images, outside tmux) and stands him under Claude's replies in the desktop app, reacting to what your AI coworker is doing. In Claude Code, /oxy replaces /hatch-pet.
  • Send private feedback from the terminal — ox feedback --file report.txt sends a report to SageOx Support and returns a receipt. --dry-run shows the exact request first, --anonymous needs no login, and the report is checked for secrets before it leaves your machine.
  • Run ox on machines that go away — ox session wait waits until a machine's finished sessions reach the Ledger and names any that did not, so a pod can be torn down safely. ox config set disposable_machine on keeps ox healthy on a machine where only the home directory persists.

Improved

  • AI coworkers get the command they meant — ox corrects more of what they type, keeps their --json choice, and only suggests a correction that would change what the command does. Commands ox itself suggests, such as ox agent <id> session log, now run, and plan list, murmur list, session list and conversation list share --limit, --since, --until and --json.
  • Drawing on a plan review is easier — drag to draw an oval, then move or resize it before you send it, and drag the review toolbar out of your way. View original now circles the words you drew around.
  • Large sessions preview, import and stop faster — previewing or importing a multi-megabyte session is up to about 4× faster, and stopping it or retrying its upload about 8× faster, with every credential check still in place.
  • Codex explains each SageOx hook before you approve it — every hook names SageOx and says what it does, including recording and publishing. ox doctor refreshes older hook labels.
  • ox query says when a discussion happened — results carry the date a meeting was recorded, not only when it was uploaded, so imported meetings land in the right place in time.

Fixed

  • Sessions in a git worktree reach the Ledger — the Claude desktop app runs each session in a worktree, which ox treated as another repository: nothing was captured, and session stop reported success anyway. Worktree sessions now record and upload.
  • ox agent <id> commands keep working — from every worktree of a repository, and for as long as the AI coworker runs, instead of failing with "instance not found" or expiring after 24 hours.
  • Resumed, compacted and forked sessions are recorded — Claude Code sessions you resume or compact record their new turns, ox session import offers resumed and forked ones, and resumed Codex sessions keep working links. The conversation a session continues is never published as part of it.
  • Sessions held on your machine reach the Ledger again — since 0.24.0, commit IDs and some ordinary session text were mistaken for credentials, and cleaning up leftover files counted as mass deletion; either one held back every session in the same Ledger commit. Held sessions publish on the next sync after you upgrade, and retrying a summary now retries its push.
  • First-time setup works without help — ox agent prime no longer hangs in tools that keep input open, a new session's first prime brings team context, first clones of the Ledger and Team Context succeed, ox team list shows your teams before anything syncs, and ox init --no-input works with a single login.
  • Your unsaved plan edits stay yours — ox doctor's repairs and session uploads no longer commit plan edits nobody saved; it finishes saves that were interrupted, and ox doctor --fix-slug=plan-unsaved-edits shares unsaved edits once you confirm. Saving an HTML plan from review or plan mode keeps the page you wrote, and a review waiting to retry survives Ledger cleanup.
  • Team skills reach your team — ox skills publish now pushes the skill and says whether it did, ox agent prime installs team skills, and ox sync skips an item this machine cannot install instead of failing. Teams that connect their Claude organization no longer see each skill twice in Claude Code.
  • Sync and repairs are more dependable — two writers syncing a Ledger at once no longer wedge it, ox's background sync stays with your repository after you upgrade instead of starting a second copy, interrupting ox doctor while it re-clones leaves nothing half-done, a new code index no longer fails when several commands open it at once, and "access denied" appears only when access is actually denied.

Security

  • Team skills can't skip approval — a team skill that grants tool access or runs a command when it loads always waits for ox skills approve, including forms ox previously missed.
  • Usage reports no longer carry what you typed — when a command fails to parse, ox sends only its command and flag names; paths, IDs, names and free text are removed on your machine first.