Vibezeo 0.8.0

Review queue.Agents that finish with changes wait in one list, across every engine, with what each one touched.

It's the Review view in the Dashboard panel ("Open review queue" in the command palette). An agent or thread joins within a second of finishing when its worktree or git folder has changes. One with no changes, outside a git repository or on an SSH host doesn't, and neither does a Read only thread, such as a second opinion's reviewer. Each item shows who did the work, on which engine and branch, its last reply or screen lines, how many files it changed, and labels for what it touches: 20 or more files, migrations, CI, lockfiles, .env files, auth paths, your public API paths and deleted tests. The labels come from fixed rules on your computer, with no AI. Sort by risk or by finish time.

In the review queue, Approve remembers exactly what you approved: if the agent changes anything after that, the item comes back as "Changed since you approved".

Open diff shows the whole change, and Last turn what changed since the agent's last checkpoint. Snooze hides an item for an hour, or until its agent finishes again. On a worktree, Send back opens its diff with review comments, Second opinion asks another engine, and Merge opens the merge confirm with its forecast. "Repository checks…" on an item sets the repository's public API paths and a check command, such as pnpm test, that runs in the agent's folder each time an item joins, for up to 10 minutes. Both are kept in Vibezeo's settings on this computer, never in the repository. The queue starts empty each time Vibezeo opens, apart from threads that had already finished.

Second opinion."Ask for a second opinion…" on a worktree has another engine read its changes and leave line comments you can edit and send on.

It's in the worktree's ⋯ menu and at the top of its diff. The reviewer is a Read only thread in the worktree, on your own engine and account; any edit it tries is denied. It gets the branch, the task if you give one, and the diff, up to 150,000 characters. Each finding on a line the diff shows becomes a comment on that line, and everything else it says goes into one General comment, so nothing is dropped. Edit or remove them like your own comments, then "Send to agent". Only engines that run as threads can review: on Free, Claude Code or Codex. One review runs per worktree at a time.

Ports and environment per worktree."Worktree setup…" can give each worktree its own ports and environment variables, so parallel agents' dev servers and databases don't collide.

With "Ports per worktree" (1 to 100), each worktree gets a block of that many ports in 41000–48999 that no other worktree on this computer has, kept until you discard it; its terminals and agents see the first as VIBEZEO_PORT and the block as VIBEZEO_PORTS. If a port in the block is busy when the worktree is made, it takes the next free block and says so. "Environment" takes one NAME=value per line for every terminal and agent in the worktree, with {{slug}}, {{index}}, {{repo}}, {{port}} and {{port+1}}… filled in, such as DATABASE_URL=postgres://localhost/app_{{slug}}. Each worktree also gets COMPOSE_PROJECT_NAME, so docker compose runs a separate stack in each, and the setup command, the teardown and Best of N's check command see the same variables. The tooltip on a worktree's row shows its ports.

The environment is trusted with the setup and teardown commands: until you trust them on this computer, terminals get only Vibezeo's own variables.

Names that change how shells start or where agents sign in or send their requests (such as PATH, LD_PRELOAD, CLAUDE_CONFIG_DIR, *_BASE_URL and proxies) aren't allowed, and nothing in the file can replace what Vibezeo sets for an agent. A value that looks like a password or key gets a warning, since the file is meant to be committed. Worktrees on SSH hosts don't get ports, environments or previews.

Previews per worktree.Run your app in every worktree at once, each on its own ports, and try each agent's version from the rail.

Add a preview command to .vibezeo/worktree.json, such as pnpm dev --port {{port}} (it needs ports per worktree). "Start preview" in a worktree's ⋯ menu runs it in a terminal named Preview, and a worktree you make with New worktree or the launch sheet starts it by itself, after its setup if it has one. Once the port answers, the worktree's row shows Preview, which opens it in a sidebar browser tab. Best of N's Compare gets Previews, which starts every agent's preview and opens a tab for each. "Compare previews…" in a worktree's ⋯ menu (or "Side by side…" in Compare) opens two worktrees' previews side by side in a new tab, each pane named after its branch. A preview runs only once you've trusted the file's commands, and each one counts toward Free's 4 live panes.

Runaway guard.When an agent may have run away, its pane shows a card saying why, and a notification goes out.

By default it trips when an agent works for 90 minutes without its turn ending, uses 30 points of its account's usage window in an hour (Codex, which reports its windows), starts 20 subagents since your last prompt (Claude Code), or ends three turns in a row with the same screen. Pause sends the agent Esc; if it's still working 5 seconds later, Vibezeo stops its program until you click Resume. An agent on an SSH host only gets the Esc, and when an agent can't be paused, the card says so. Dismiss keeps that limit quiet until your next prompt. Settings › Teammates › Runaway guard changes the limits, turns any of them off, or has the guard pause the agent itself. Threads and teammate chats are watched for working time only, with a notification. The guard never sends a prompt and never switches accounts.

Usage forecast.For engines that report usage windows (Codex does; Claude Code doesn't), the usage chip in the title bar forecasts whether each window runs out before it resets.

From the last hour's readings, each window in its popover says "Not enough data yet", "On track" or "Runs out at 3:40 PM, 30 min before it resets". "Who's using it" lists the sessions that spent each Claude Code or Codex account's tokens in the last hour, with their shares, named after the pane running each one and read from the engines' own logs on your computer (token counts only). "Warn me before usage runs out" in Settings › Notifications (off by default) sends one notification when a window is forecast to run out before it resets, and the launch sheet says when an agent's account is at 80% or more or forecast to run out. Vibezeo never switches accounts for you.

One instructions file for every engine."Instructions…" in a local folder's menu in the rail shows which instructions file each installed engine reads, whether git tracks it, and how it differs from AGENTS.md.

"Use one file for all" makes AGENTS.md the source (with none yet, you pick the file it starts from): each installed engine that needs it gets a one-line file holding @AGENTS.md (CLAUDE.md for Claude Code, GEMINI.md for Gemini CLI), while Codex, Copilot CLI, Cursor Agent and Antigravity CLI read AGENTS.md themselves. Engines whose files haven't been checked show as Unknown and get nothing. You see every file it would create or replace before anything is written, and Undo puts them back as they were. "Copy into worktrees" adds a personal file git ignores, such as CLAUDE.local.md, to the worktree setup's copy list. Vibezeo never commits these files.

Antigravity CLI and Kiro CLI.Two more coding agent CLIs join the launch grid, so Vibezeo now detects 13.

They run in terminal panes only, as their vendors ship them, never as threads or in the background. Each Kiro CLI account signs in inside its own home (Free has one account per engine); Antigravity CLI keeps one account per computer, because it stores its sign-in in the system keyring. Their panes show when they're working or done, but not yet when they're waiting for you. Gemini CLI stays listed, with a note in Settings › Engines and the launch grid's not-installed list: Google serves it to paid API keys and Gemini Code Assist Standard or Enterprise, and personal Google accounts use Antigravity CLI.