Vibezeo 0.13.0

Claude Code threads summarise their edits too. A finished Claude Code turn says which files its edits changed, a subagent's and a notebook's included, with the lines added and removed where the edit says them, and each edit's row shows its own counts. When an edit's lines aren't known, such as a notebook's, a file written over whole or a file an MCP server's tool wrote, the turn lists its files without line counts rather than a partial total. An edit that was refused or failed counts nothing, a subagent's edits count even when its task ends in an error, and files changed by shell commands aren't counted.

The rail's marks say it in a word. A worktree's pull request checks read as the status bar has them, CI running, CI passed, CI failed or CI cancelled, and hovering says which checks failed. The overlap mark shows the symbol two worktrees share, such as applyCoupon, and a screen reader hears those words first. A row that's short of room shows a mark's icon rather than cut its words, and a pane header's branch gives way before the pane's title does.

The rail never cuts a row off. A row too narrow for everything it shows folds it: a worktree's Preview to its globe, its marks to their icons, and last its agents' status to a dot that still says Needs you or 2 working when you hover it. The rail stops at the width of a fully folded row, about 250 px rather than 180, a little wider while a row shows a jj badge or a preview, and widens itself to fit one. A row's ⋯ shows in place of its agents' count, never over a mark.

Checks follow a push. After you push from a worktree's git panel, its pull request's checks are read at once and every 2 minutes for the next 5, so its row moves to CI running when GitHub starts the new run instead of up to 15 minutes later.

An account is named when there's a choice. A pane's header and its Dashboard card name the agent's account only when its engine has more than one, so a single account's "Default" no longer shows.

A new agent no longer looks busy before its first prompt. An engine that reported starting before Vibezeo read its banner could show "working" on the Dashboard, and to Auto, until you typed to it. Its report of starting now counts, whichever arrives first.

Memory and skills save what their own check allows. Their editors read the whole text, so Bearer at the end of one line with a value on the next looked like a key, and Save stayed off for text that saving accepts. They now read a line at a time, as saving does.

A finished chat stays unread until you look at it. A chat you'd opened earlier and left could lose its unread dot when its turn ended in the background, and one you were looking at could get the dot back. Now a chat is marked read only while Thread mode shows it (the selected chat, or a chat in the grid) with Settings closed and Vibezeo's window in front and focused: when you open it, when its turn ends, or when you come back to Thread mode or to the window.

Allow can't move under your click. A chat's Allow and Allow for this thread, held actions' Approve & run and the review queue's Approve now wait until they've held still on screen for half a second, and while anything covers them, like the Dashboard's and the pane's Allow. So the next request sliding into place, or a new held draft arriving on top, can't take a click or a double-click meant for something else. A chat card's keys answer only the card you're on, and only once its Allow would. A phone's queued Allow in Settings › Remote, Trust and connect for a new SSH host key and Trust and run for a repository's worktree setup wait the same way. A waiting Allow can still be reached with Tab, and a card scrolled out of view takes no key.

The task list keeps your latest changes. When changes came quickly, an earlier refresh that finished last could show the list as it was before them until the next change.

A plan's runs show where they are. When a plan started, a run that finished while Plan with Lead was first reading the runs could stay shown as in setup or running, with no verdict, until you opened the plan again.

A plan's counts read right for one. A plan says "1 with changes requested", where it said "1 changes requested".

A Dashboard card's line follows its agent. While Claude Code works, its card names the tool it's running, and the name goes when it stops. Before, an open Dashboard named no tool, and a card that opened mid-tool kept that tool's name after Claude Code had finished.

A Dashboard card keeps up while the cards load. An agent that changed while the Dashboard was reading its cards could show its earlier state until it changed again, and a new agent's card could go missing until the agent next reported.

Screen readers name the rail's unread dots. The dot on a workspace or a pane that finished while you were away is read as "Finished while you were away". Before, screen readers skipped its label.

Text from elsewhere can't hand an agent a file or run a command. What Vibezeo passes an agent from somewhere else can't make Claude Code read a file on your computer: a worktree's diff and a second opinion's words, in the review you send or for the second opinion itself, a pull request's review comments, a failed check's log, Warn both agents' note, a teammate's mail, the terminal selection, clipboard, file name or branch name in a prompt, a branch's name in a review you send back, a web page's elements, the turn a teammate learns from, a teammate's briefing for an ACP agent, a handoff brief, the summary that continues work on another account, and what Auto and Conductor send. Any @ there that could name a file is written ＠, and the message says so. Characters nobody can see, and any that Claude Code or typing would drop, go first, so they can't hide a word or a command. A paste marker such as [Pasted text #1] can't stand in for something you pasted, words such as ultrathink get a hyphen, and a word such as /release or /apps/web:release, which could run one of your skills, gets a full-width slash. A handoff brief, or what Auto or Conductor sends an agent, that starts with / or !, or is just "exit" or "quit", has its first character written full width (／, ！), so it runs no slash or shell command and quits nothing, mid-turn too. What Conductor types in parts, without Enter between them, can't join into such a word either.

A key split by a character nobody can see is still hidden. Vibezeo hides keys and tokens before text reaches a model, your phone or anything else off your computer, and now also finds one that an invisible character splits, sets apart from the word before it, or both: in what Conductor reads from panes and threads and what it says, the review queue's lines, a plugin's replies and errors, a teammate's brief and memory, what your phone shows, crash reports, usage statistics and Copy diagnostics. A plugin's replies now lose every key Vibezeo recognises, not only the connection's own. A thread that switches engines hides keys in the summary it hands the next engine too, as a handoff's brief does.