Changelog
What changed in each Shipyard release, newest first.
Releases are on GitHub.
The Mac app updates itself; a box
upgrades with berth upgrade BOX, or by running the
install command
again.
Unreleased
- Boxes are ready for the Shipyard phone app. With phone access on, a
box now also lets a paired phone read an agent's recent conversation
(never tool output or files), follow which agents and runs change as it
happens, stop an agent, and ask to be notified through Shipyard's push
relay instead of ntfy. Notifications still never carry what an agent
wrote, a new code forgets every phone, and
before:phone.pushhooks can refuse a phone. See the phone API. - Quiet projects are grayed out. On the rail and in the two-column sidebar, a project with no agent or shell open in it is grayed out, so the ones you're using stand out. Switching to one of those opens New task in it; a project whose agents are only idle or done just switches to where you were, as before.
- Agents come back after a reboot. Starting a chat agent again after
its box restarted ran its old command with its last word cut off, so
Claude refused to start ("argument missing") and the tab only said it had
ended. It now carries the conversation on with Claude's whole command and
--resume. An agent that exits as it starts says it couldn't start, with what it said, instead of looking resumable, and berthd no longer stops following a conversation that is started again on its own record. Another agent working in the same folder never makes a conversation read as running. When berthd starts, agents it remembered by folder that nothing runs any more show as ended, and folders that are gone are forgotten. - Worktrees fits a narrow window. With the sidebar open on a small window, the table and its filters ran off the right edge. Now the Port column steps aside first and "vs base" after it (both are in a worktree's tooltip), the show-only filters fold into one Show menu, and the last commit keeps room to be read. Kits' cards start under the page's title and line up their footers, and icon buttons give a little when you press them (not with reduced motion).
- Home opens on what needs you. The header picture is now a band, so Needs you and Working now show without scrolling, even in a small window. Every agent row says which project and worktree it is in ("shop / checkout-fix · devl") before what it is running, and Customize moved under the widgets. Before your first project, Home shows only the way to add one. In the command palette, rows under Needs you no longer say it again, so their titles fit; in the notification centre, titles under Needs you drop "needs you", each row says where first and then which agent, and the panel opens below the top bar instead of over it.
- Review is easier to read with many projects. Work waiting for review is grouped under each project and its box, and J and K still walk from one group to the next. With one project, each row says where it is, project first ("shop / checkout-fix · devl"). Approve and Send back lead the actions; Open, Mark reviewed and a ⋯ menu with Discard sit apart on the right, and A, S, D, E and Enter work as before. The agent's last message reads as wrapped prose instead of the terminal's broken lines.
- Upgrading a box no longer leaves dead processes behind. Each upgrade left the terminals that were open on the box, and any command berthd had running, as zombie processes under berthd until it restarted for good. A box now closes its terminals before it switches to the new build, and the new build cleans up whatever the old one left. Your sessions keep running, and the app reattaches its terminals as before.
- A box with many projects sits idle again. berthd ran three or four git commands per project every time the app, a plugin, the MCP or its own worktree watcher listed them, dozens a second on a box with a few dozen projects. It now keeps each project's worktrees, remote and default branch, and asks git again only when the repository changes (or after a minute).
v0.3.31
-
Worktrees with submodules can be removed. git refuses to remove any worktree with a submodule checked out, even a clean one, so removing or archiving one failed. Shipyard now checks the worktree and its submodules itself and removes it when nothing would be lost. If a submodule has changes, stashes, or commits that aren't pushed anywhere (a commit only a local tag points at counts), the worktree is kept and you can still choose Remove anyway.
-
Pin worktrees to the top of the sidebar. Right-click a worktree, or open its ⋯ menu, and choose Pin to top. Pinned worktrees sit in their own list above your projects, in the order you pinned them, whether or not anything runs in them. Unpin puts one back under its project. Pins are kept on this computer, and a worktree's pin goes when you archive or remove it.
-
Switching projects no longer opens a new task. Clicking a project on the rail (or in the projects column) with no agent at work in it now takes you to the worktree of it you were in last, or its main checkout, instead of popping up New task.
-
The app has a new icon: a box in its tray with the amber lamp lit on top. A dev build wears the same box on orange.
-
A structured agent's questions can always be answered from the chat. An agent that asked you something while helpers it left in the background were still working showed a bare "Claude asks you" card with no options, and a reply box that took nothing. Now the chat shows the agent's form with its options, however much its helpers wrote after the question, and a helper reporting back no longer counts as your answer. A card with nothing to pick says Reply below to answer it, and what you type there is the answer.
-
A chat no longer goes blank at the top mid-turn. When the agent's long reply folded into Working and the chat got short enough to fit, the Mac app could leave your prompt and the first steps undrawn, with empty space where they belonged. They now stay drawn.
-
An agent waiting on its own work no longer looks like a live turn. When turn check reads a finished turn as work still running (a build, a deploy, a task it waits on), the sidebar and its tab show an hourglass that says so, instead of the spinner of a turn still going. Turn check also reads a handoff as done: an agent that started a worker it won't follow and told you that you can close the session no longer shows as working.
-
Plugins can give an agent a file.
berth.api.attachuploads a pasted or dropped file into a session's worktree, or a worktree an agent is about to start in, the way the composer does, and resolves with its path on the box to put in a prompt. A pasted screenshot shrinks first, and a path to a file on this computer (berth.api.localPathsfinds them in pasted text) goes up through Shipyard. Files up to 20 MB. -
A chat whose turn ended no longer looks busy. When an agent's turn ends while its work runs on (turn check read a build, a deploy or a delegated task as still going), its chat now ends on an hourglass, Waiting on its work, instead of a spinner and a ticking timer, and the reply box shows Send rather than Stop: what you type starts a new turn.
-
Turn check reads a status report on another session as done. A reply that says how another session or worker is getting on and offers to follow it up ("I can nudge it, or tell you when its turn ends") no longer keeps the session showing as working. Waiting on another session still counts as working when the agent says it will act once that session reports back.
-
A helper working in the background no longer restarts a finished turn. Tool use by a Claude Code subagent left running after the agent's turn ended used to put the session back to running until your next prompt.
-
Stopping several agents can archive their worktrees too. The Agent Dashboard's Stop and Clean up dialogs have a box, Archive their worktrees where it's safe. Ticked, each stopped agent's worktree is archived once its agents are stopped: the archive script runs and the branch is kept. A worktree with uncommitted changes, or with another agent still running in it, is kept as it is, and its row says why. The dialog remembers whether you ticked it.
-
Archiving never tears down uncommitted work. A box now refuses to archive or remove a worktree with uncommitted or untracked files before it stops anything or runs the archive script, instead of running the script first and only then finding that git won't remove it. Remove anyway still forces it.
-
Cards for agents in chat say what the agent is doing. On the Agent Dashboard, an agent that runs in chat showed the frames of the host that runs it ("out stream_event") and "stdio" as its prompt. Its card now shows what it asks while it waits (the command and why, or the question) with its answers, otherwise the last lines of its reply, and the prompt it started with. Cards for agents in a terminal are as before.
-
Shortcuts show your keyboard's keys. Back and Forward are ⌘ and the two keys right of P. On a layout where [ and ] need ⌥, the arrows' tooltips, the Keyboard shortcuts sheet, Settings and ⌘K now show the keys in that place, such as ⌘Å and ⌘¨ on a Norwegian keyboard, instead of ⌘[ and ⌘]. A US layout still reads ⌘[ and ⌘].
-
Shorter shortcut descriptions. Each fits on its line in the Keyboard shortcuts sheet, without the "…". The details they had are in the new Keyboard shortcuts guide.
-
Back and Forward's tooltips show on hover. Once an arrow had somewhere to go, hovering it showed nothing; only the keyboard's focus did.
v0.3.30
- Agents on a Mac run at full speed. Shipyard's login agents (the
Mac app's agent, and
berthdon a Mac box) used to ask macOS to run them as background work, which slows their CPU and disk, and with them every agent and build they start. They now run as interactive work. Updating the app applies it to this Mac right away. A Mac box updated with Update all gets it the next time it restarts or you log in there; running the install command again applies it at once. - Back and Forward, as in a browser. The arrows at the top of the sidebar, ⌘[ and ⌘] (also in the Go menu and ⌘K), and a mouse's back and forward buttons take you through the places you've been: a page such as the Agent Dashboard, or a worktree and the tab you had open in it.
- Archiving the worktree you're in takes you back. When the worktree in front is archived or removed, you land on the place you were before it, such as the Agent Dashboard, instead of Home. The dashboard also keeps where you'd scrolled to.
- The + next to Worktrees makes a worktree. In the project sheet beside the rail, it now opens New worktree for that project, instead of a menu to add a project or a box. Those stay on the rail's own + and the boxes row at the bottom of the sidebar.
v0.3.29
-
Removing a worktree no longer asks first. Remove worktree in the sidebar, Stop and remove worktree on the dashboard, and Delete in the Worktrees view now act at once: the worktree shows Removing… and goes when its box gets to it, one at a time. Its branch is kept. One with uncommitted changes stays, and the note that says so has Remove anyway.
-
Switching projects takes you to where the work is. Picking another project in the sidebar now opens the agent in it that worked last, one that's working or waiting on you. With none, the new task box for that project opens instead.
-
Answers from the question card reach Claude Code, long ones too. When Claude asked several questions with lists in them, or you wrote a long answer, the card's answers often stopped before the first one and left Claude's own form untouched: an agent nobody has open has a screen too short for the form. Now the box makes room while it answers, then puts the screen back.
-
Answers that don't go are kept. If the box still can't fill in Claude's form, the card keeps every answer and says why, with Retry and Copy answers, and Claude's own screen opens under it to finish there.
-
⌘-click a Markdown path in a terminal to read it here. A
.mdfile of the worktree, such as one an agent names in its reply, opens in a File tab as a page. ⌘⇧-click sends it to your editor, as ⌘-click does with every other file. -
Pictures show in a Markdown file's Preview again. Since v0.3.27 a
.mdfile's pictures, from the worktree or the web, didn't show. -
A new Claude chat opens in Claude Code's terminal before you've said anything. Switching one to its terminal used to flash "No conversation found" and end the session. Now the terminal opens on the same conversation, and the chat picks it up again when you switch back.
-
No made-up question card. An agent whose last message asks you something still shows as Needs you, but its chat no longer adds a "Needs your answer" card with Yes and No that it can't answer with (the card took its title from Claude Code's footer, often "/rc"), and the Agent Dashboard no longer offers a numbered plan in that message as options to pick: answer it from the reply box. Options show only when the agent's own screen offers them, and the mode menu works while such an agent waits.
v0.3.28
- Every worktree shows in the sidebar. Worktrees with nothing going on are no longer folded under "Show N older", so the sidebar's room goes to your worktrees. The Active / All worktrees choice went with it. A project still folds from its header, and a worktree's handed-off children from their count.
v0.3.27
Keeping main up to date and opening Codex edits need each box's new
berthd: use Update all in the status bar, or berth upgrade BOX.
-
Main stays up to date. A new worktree from
mainstarts from whatoriginhas now, not from whenever the box last pulled: the box fetchesmainfirst and fast-forwards your local one, so the main checkout and diffs againstmainare current too. It never moves past uncommitted changes or commits of your own. Turn it off per project under Main branch in Project settings. -
Markdown files open as a page. A
.mdFile tab shows the rendered page, its source, or both side by side, from three buttons in its header. Links to other files open them in a tab, and the worktree's pictures show. Pictures from the web, such as a README's badges, now load in the app. -
Codex edits open again. In a chat with Codex, opening an edit's This change (or a command's output) said "that step is no longer in the conversation". It now shows the change. Edits by Claude keep their line numbers too.
v0.3.26
- The sidebar is a rail of your projects. Each project is a tile on a narrow rail, with its GitHub owner's picture or a pattern made from its name, and a dot that turns amber when an agent in it needs you and blue while one works. Pick a tile to see that project's worktrees beside it, with search and New task at the top. Review, Automations and your other places sit at the top of the rail, and Settings at its foot. With one project there's no rail. Settings › Appearance › Sidebar switches back to the single list or the two columns.
- Everything lines up. The app is laid out on one grid. The bar across the top is one height on every screen (the sidebar's strip, the tabs, Review and the other full pages), with the window's buttons and everything else in it on one centre line. Every pane's toolbar is one height too. In each sidebar layout, rows, their icons and labels, and their right ends sit on the same lines, and each full page's content starts right under its title. Full pages now title themselves in the top bar, at the size of a window's title.
- Agents run in chat, with nothing to turn on. New Claude Code, Codex, OpenCode, Gemini CLI, Cursor Agent and oh-my-pi agents open as a conversation that the box drives directly, whether Labs is on or off. The Structured agents switch is gone from Settings → Labs, and Permissions (Bypass or Ask) moved to Settings → Agents and models. Agents that other agents, the CLI, runs and automations start run in chat too, asking for each permission on a box Shipyard hasn't reached yet. Claude Code's own terminal is still one click away in its pane, and a box too old for chat keeps running agents in their terminals until you upgrade it. Agents in chat has the details.
- Attempts on a repository GitHub doesn't host finish with your pick.
On Gitea, Forgejo, GitLab or a folder, picking an attempt with a draft PR
turned on failed the run at
gh pr createand left every worktree behind. Shipyard now pushes the picked branch, says plainly that it didn't open a pull request there (with the forge's own link to open one when it prints it), archives the other attempts and finishes the run. - A pull request that doesn't open no longer fails the run. When
ghcan't open the pick's pull request (a sign-in that ran out, say), the run waits with your pick kept and nothing archived, and says it is publishing that failed, not the agents' work. Review's Compare and Automations → Runs offer Retry publishing or Finish without a pull request;berth run approveandberth run rejectdo the same. - A saved Compare tab no longer leaves the window empty. Quitting with a Compare tab open left the next launch, and every one after, on a dark grey window with no sidebar, until the saved tab was removed by hand. It now opens as it was. Saved tabs that can't be drawn are left out rather than stopping the app, and a pane that breaks says so in its own place, with a way out (its terminal, or closing the tab), while the sidebar and your other tabs carry on.
- A helper that is back shows as back in a chat session. In a chat (structured) Claude session, a helper Claude left working in the background stayed "working" in Crew, its time running on, after it had come back: the stream says so in a line of its own that wasn't read. It now shows back, its time stopped when it returned.
- A turn Claude starts by itself shows as working. When a helper (subagent) Claude left in the background came back, Claude went on in a turn no prompt started, and the session showed as done while it worked: no spinner and no Stop. That turn now shows as working until it ends.
- A helper's permission no longer gets lost. When Claude Code started helpers (subagents) in a chat session and they asked to run something while Claude asked too, or after its turn ended, the app showed the newest ask while Yes answered the oldest, moved on to "working" after one answer, and dropped the helpers' asks when the turn ended: they waited forever. Each ask now shows in turn, the one shown is the one you answer, and a helper's ask stays until you answer it or the helper is stopped.
- Another Claude in the same folder no longer stalls a session. A Claude
that Shipyard didn't start (in an editor, or a one-off
claude -pa tool runs) in a session's folder could mark the session's agent exited when it quit; the session then showed as working for good and Stop did nothing. Its events are now told apart by its own conversation. - A two-column sidebar. Pick it in Settings › Appearance › Sidebar and your projects sit in one column, with the tasks of the one you pick in the next. A project's row turns amber when something in it needs you and blue while something works. The tasks column can be resized or folded away, and opening a worktree shows its project.
- The sidebar is quieter. Its text is one colour in a few strengths, and colour means only two things: amber when an agent needs you, blue while one works. A finished task gets a quiet check, an idle one no mark at all, and times show only on tasks that are doing something.
- Boxes stay out of the way until they matter. They no longer have rows in the list: a project on a box that's away dims and says "offline", and the line at the foot of the sidebar says which boxes are up. Its tooltip has each box's load, which leaves the status bar. A box's actions are in its projects' menus and in the foot's menu.
- Text is the same size for the same thing on every page. Chats, Settings, onboarding, Home, Review and the rest use one set of sizes, and ⌘+ makes all of them bigger.
- A queued prompt that failed is amber in the status bar, like anything else that needs you, rather than red.
v0.3.25
- The sidebar is a size smaller. Its rows are the size of Settings' list: 13px names with small icons, project names as quiet labels, and the same height for tasks, the worktrees under them and your boxes.
- Boxes that are offline or slow sit at the foot of the sidebar, just above the boxes' line, so your projects stay at the top.
- A worktree's menu is shorter. It is four groups: what to start in it; where to open it and what to run; its name, colour and what to copy (path, branch, URL and review links are under Copy); then stop, archive and remove. Resuming a conversation is under New agent, and Compare and tabs beside are under Open in.
- A chat that recovers from a read error no longer crashes the app. A pane that went from "Couldn't read the conversation" to its chat drew with one hook more than before and showed "Shipyard hit an error".
v0.3.24
- Sidebar rows hover cleanly. A row stays lit while the pointer is on its
- and ⋯, instead of the buttons sitting on a patch of their own, and folded projects sit one under another without big gaps.
- The sidebar lists your tasks. Each row names what its agent is working on, with a dot, spinner or check for how it's going and how long it has been that way. Worktrees an agent handed work to sit under it on a line, a count beside it folds them. Home is called New task, notifications and hiding the sidebar sit at its top, your boxes at its foot, and notices are one quiet line each. The project menu moved into More.
- Settings is a full page now. Its sections are grouped in the sidebar's place (Esc goes back), each page is one calm column of settings, and Find a setting (⌘F) jumps to any of them. Appearance picks Light, Dark or Match system above the themes.
- A plugin's sidebar item can show a count.
addSidebarItemtakesuseCount, shaped like a Home widget's: an amber count on the item's row, and on More's while the item is under More. Sessions waiting in a plugin's own place, such as a project kept off the Agent Dashboard, now have somewhere in the menu to show. AuseCountthat throws shows no count and leaves the sidebar alone. See Sidebar counts. - Codex sessions get their own status again with Codex's shared daemon.
Codex can run every session from one background daemon, which runs hooks
with the first session's environment: with two Codex sessions on a box,
the second one's prompts, tools and turn ends showed on the first (which
was even renamed after them), and the second only ever read off its
screen. Shipyard now starts Codex with
--no-daemonwhen the installed Codex knows the flag, and a Codex hook from another session's folder, or about a conversation another session holds, goes to that session. - Every page uses the same type sizes as Settings. Review, Worktrees, Automations, the Agent Dashboard, Kits, Project settings, Team setup, Home, dialogs and sheets now share one set of sizes: a large page title with a quiet subtitle, bold section titles, medium item names, 13px descriptions and slightly larger metadata and code, so the app reads as one piece.
- A box with a full hostname is named by its first label. A cloud VM
whose hostname is
devbox.europe-north1-a.c.example.internalcalled itself by all of it: the phone and notifications showed the whole name, and it can't be one label of a worktree's URL, so$BERTH_URLwas empty andberthd browser openrefused every worktree. A box (and a box a laptop pairs) is now named by its hostname's first label,devbox, with_becoming-. Restart berthd on such a box to pick up the new name; existing laptop pairings keep theirs. - Turn check: a check-in an agent scheduled for itself is Idle. An agent that ends its turn saying it will check again at a later time ("tomorrow morning I'll check the numbers"), with nothing running now, shows as Idle, even when the machine is kept awake for that check. It used to show as Working. Work still running in the background, a re-check or a watcher that reports back, stays Working. Sessions and agents
v0.3.23
- Project settings opens for a project with a kit's flows. A kit whose flow used a step the flow editor doesn't edit (a PR, a loop, a gate) made Project settings show "Something broke" every time. That flow now shows in Automations with a plain icon.
- A long Codex turn stays working until it ends. Codex's
notifycomes at the end of each of its messages, not only the turn's, so a turn read as finished after its first message. Once a session's Codex hooks have said a prompt started, only itsStophook ends the turn; Codex versions without hooks still finish withnotify.berthd integrations install codex(and an upgrade) also addsPreToolUseandPostToolUsehooks: a question in Codex's question box shows as needs you, and answering it, or Codex carrying on by itself after a turn ended, shows it working again. If Codex asks, trust the new hooks with/hooks. - An agent asking you a question no longer reads as done. When a task's first prompt led Claude Code to ask something (its question form, with choices), Shipyard could read the question off the screen as nothing, end the wait and mark the agent finished while it waited for you. Now a wait Claude's hooks report lasts until Claude goes on, the turn ends at Claude's own Stop, and the screen recognises the question form too. Sessions and agents
- No more tall blinking bar at a terminal's left edge. With the Ghostty renderer, a focused terminal showed the browser's own text cursor down its whole height, next to the terminal's real cursor. Only the real one shows now.
- A box newer than the app is no longer called outdated. An older
Shipyard used to offer to update a box that already ran a newer release,
and Update all (or updating boxes automatically) would put its older
berthdon it. Now a newer box is left alone, andberth upgraderefuses to downgrade one unless you pass--force. This works once both the app and the box are on this release or later. - Keep a project off the Agent Dashboard. In Settings → Agents → Agent Dashboard, pick projects whose agents you follow somewhere else, such as on a plugin's own board. Their agents leave the dashboard and the waiting and working counts in the sidebar and status bar, and the dashboard's header says how many projects are kept off. They still show in the sidebar, on Home and in notifications. Show brings a project back.
- Project changes made quickly all stick. Renaming a project, moving it between sections or keeping it off the dashboard twice in a row could lose the first change, because both saves read the same version of your projects. They now save one after the other.
- A kit link pinned to a commit works. A GitHub link copied from a
commit's page (
…/tree/308bc75…/kits/web) used to fail with "Git couldn't do that", because git can't clone a commit by its hash as if it were a branch. Shipyard now fetches the kit at that commit.
v0.3.22
- Mac terminal keys work in terminals. ⇧↵ starts a new line in Claude Code instead of sending, ⌥← and ⌥→ move a word, ⌘← and ⌘→ go to the start and end of the line, and ⌘⌫ deletes back to its start. ⇧↵ needs the box upgraded; a session's agent started before then picks it up when it restarts. Terminals
/resumeworks in a structured agent's chat. It opens a list of the agent's past conversations in that folder, newest first, including the ones you had in Claude Code's own terminal. Pick one and it carries on in the same tab, with what was said before shown in the chat. The conversation you leave stays under Resume a conversation. ACP agents that can list their sessions (OpenCode, for one) get the same list; Codex's isn't there yet. Needs the box upgraded.- A skill Claude loads shows as one step. In a structured chat, the skill's whole text used to show up as a message from you. Now the chat says "Used the berth-orchestrate skill" among the agent's steps. Needs the box upgraded.
- New terminals no longer open with a systemd warning. On boxes with systemd 258 or later, every new tab started with D-Bus property CPUAccounting is deprecated. Sessions still report their CPU use as before. Needs the box upgraded.
v0.3.21
-
The phone app pairs from a pasted link. Its Pair and Boxes screens have Or paste a pairing link, for a phone app added to the home screen, which keeps its own storage apart from the browser. A good pairing link no longer says it is damaged, and Back always returns to the list instead of leaving the app. Needs the box upgraded. Phone
-
Turn check: a turn that ends on a question needs you. Off by default; turned on for a box in Settings → Agents, berthd asks a decision model you pick about each finished Claude Code or Codex turn: Jev on Vercel AI Gateway, OpenAI Decisions (
gpt-6-luna), or Laya on a server you run, which keeps turns on your machines. One that asks you something or is blocked shows as Needs you; one that stopped while something it started still runs (a build, a deploy, a watcher) as Working; and one parked until someone else replies or a later time ("check again in 24 hours"), with nothing running, as Idle. Neither of those notifies you. A failed or unsure answer, or work that is done, leaves it Done. It sends the turn's prompt (its first 4,000 characters, looked for further back when a long turn's tool calls pushed it out of view), the end of its reply and the session's title, with anything that looks like a secret taken out first. It is set only from a paired laptop, never from the box's own socket; each provider's key is kept apart and goes only to that provider. -
Switchboard: Pick up focused and recent conversations and worktrees with ⌘⇧K or from Home. Type to find one, ↑ ↓ and ↵ to open it, ⌘S to star it into Focus, ⌘E to leave a pickup note, and Tab to step through your areas. Right-click a row (or ⇧F10) to move it to another area. Your focus, areas and notes stay on this computer.
-
Choose the picture at the top of a new chat. Settings › Appearance › Header picture picks the painting across the top of a new chat, an empty worktree and Home: the harbour (still the default), one of six new paintings (Alpine lake, Meadow, Desert mesa, Snowy fjord, Lavender fields, Sea of clouds), a picture of your own, or none. Each new painting has its own night version for dark themes, as the harbour does, and a picture of your own is dimmed there.
-
Six new chat backgrounds. The same six paintings are scenes under Settings › Appearance › Chat background, at night in a dark theme.
-
Math in code blocks draws too. A
```mathblock, as GitHub writes a formula, and a```latexor```texblock of equations now show as typeset math. A whole LaTeX document (with\documentclass, meant forpdflatex) stays code, as does a block still being written. -
Agents started by other agents follow Structured agents too. With Structured agents on in Settings › Labs, an agent that another agent starts with the berth tools, one started with
berth task neworberth session new, and the agents of runs and automations now run structured, with your Permissions choice, instead of always in their terminals. Update your boxes (berth upgrade BOX, or Update all) for it to take effect. -
Mermaid flowcharts in a reply draw as diagrams. A
mermaidcode block with a flowchart shows as the diagram, in the theme's colours, the way a diagram artifact does; Source in its corner shows what it was drawn from, and Copy still copies that. Other Mermaid diagrams, such as sequence diagrams, show as written, and a reply still being written shows its blocks as text until it's done. -
Mermaid flowcharts draw as they were written. Subgraphs show as titled, tinted boxes around their nodes, nested where they nest, and a
<br/>in a label starts a new line. Labels lose their quotes, codes such as#quot;and&show as the characters they stand for, and nodes keep their shapes: pills, cylinders, circles, hexagons and slanted boxes. Every kind of arrow is read, labelled or not, as are chains likea --> b --> canda & b --> c. A loop back up the chart curves round the side instead of stretching it. This goes for diagram artifacts too. -
Diffs in a reply look like the chat's edits. A
difforpatchcode block draws with file names, line numbers and highlighting, as an agent's edits do. Plain lines of+and-show coloured, without numbers. A block that isn't really a diff stays as written.
v0.3.20
-
Pictures from a worktree show in replies. An agent can include a Markdown image such as
and the chat shows a thumbnail after checking it belongs to that worktree. Click the image to open its File tab, or use Source to see the Markdown. External image addresses stay as text instead of loading in the app. -
Tables in a reply can be explored. An agent's Markdown table gets an Explore view with sortable columns and a failures filter for status results. A small numeric table can switch to a chart in Shipyard's theme. Original brings back the table exactly as written.
-
Test failures in a reply have a short result card. Vitest, Playwright and Go test output in a code block shows pass and fail counts and the failed tests. A file and line in the worktree opens from the card; Source shows the full output.
-
Updating lands on the newest release. An update the app had already downloaded used to wait for your restart even after a newer one came out, so you went one version at a time. Now a newer release replaces the one waiting, and Restart to update checks once more before it installs.
-
A new agent no longer flashes "has ended" as it opens. For a moment after it started, its chat showed Claude Code has ended with Start Claude Code again, before the box had listed the new session.
-
Callouts in a reply look like GitHub's. A quote that starts with
[!NOTE],[!TIP],[!IMPORTANT],[!WARNING]or[!CAUTION]shows as a tinted box with its icon and name, in your theme's colours. -
Issues, pull requests and commits in a reply are links. In a chat whose repository is on GitHub,
#123,acme/shop#45and a commit's hash open on GitHub when you click them. Colours like#fff, line anchors like#L14, code blocks and existing links stay as they are. -
Files an agent names link to them. A path of the worktree in a reply's code, like
src/app.ts:14orsrc/app.ts:14-30, gets a faint dotted underline. Hovering it shows those lines, drawn like the chat's edits; clicking opens the file at that line (⌥-click beside the chat, ⌘-click in your editor). Commands and names that aren't files stay as they were. -
Math in the chat. Replies with LaTeX (
$$…$$,\(…\)and\[…\]) show as typeset math, centred on its own lines for a display. A single$is left alone, so "costs $5 and $10" reads as it was written, and math in code stays as code. A formula that doesn't parse shows as its source, in red.
v0.3.19
- Labs: a structured agent keeps its permission mode when it's resumed. A mode chosen while the agent was working could be lost, so a Codex agent set to Bypass permissions came back asking again once resumed.
- Labs: a structured agent's chat no longer opens a terminal under it. When the agent was idle, "Claude Code is showing its own screen — answer it here" opened with the log of the process running the agent, and typing went into that log. Now it opens only for agents running in their terminal. For a structured Claude Code, Show terminal opens Claude Code's own terminal on the same conversation, as the pane's terminal toggle does. Structured Codex and ACP agents have no terminal of their own, so they don't offer one.
v0.3.18
- Labs: every way of starting an agent starts it structured. With Structured agents on, only the composer started agents structured; the sidebar's and a worktree's New agent, the launcher, the new-tab and split menus, the worktree picker and hand-offs (and reviews) started them in their terminal. They all start structured now.
- oh-my-pi is a built-in agent. Start it from a project like any other
(
omp), with a model and thinking level. With Structured agents on, it runs through the Agent Client Protocol: its replies and tool calls fill the chat and its permission requests come to it.berthd integrations install ompadds a small extension to~/.omp/agent/extensionsso a terminal session says when it is working, needs you or is done, and adds Shipyard's MCP server to~/.omp/agent/mcp.jsonso omp can drive other sessions; a box that has omp gets both with the other agents' hooks. Needs the box upgraded. - Labs: a structured agent's first permission waits for you. When a structured agent asked for a permission in its very first turn, with Permissions: Ask, the box read the session's screen, which shows nothing of it, ended the wait a second later and then the turn: the chat had no Allow / Deny, said the agent was done, and the agent waited forever. The screen no longer settles structured sessions; their own stream does. Needs the box upgraded.
- Settings › Boxes reads correctly to a screen reader. A box with no browsers running put its "No browsers running" note inside the browsers list, so assistive tech announced an empty list with a stray line in it. The note now stands on its own.
v0.3.17
-
Agents sharing a worktree each get a browser tab of their own. When several agents check pages in one worktree at once, each passes
--tab NAMEtoberthd browserand gets its own page, cookies and size, so they no longer take the page from each other. Every tab keeps rendering at full speed while nobody watches it. The agent's view in a browser pane shows the tabs, stays on the one you pick, and no longer jumps from one agent's page to the next. Needs the box upgraded. -
Labs: OpenCode, Gemini CLI and Cursor Agent run structured too. With Structured agents on, they start through the Agent Client Protocol: their replies stream into the chat, their permission requests come to it (or, with Bypass, the box allows them), Stop cancels their turn, their own modes are in the chat's mode menu, and they can be resumed. A repository's own agent can join in by giving its preset an
acp. Needs the box upgraded. -
Labs: Codex agents can run structured too. With Structured agents on (Settings › Labs), a new Codex agent runs as Codex's app server, driven by the box: replies stream into the chat, commands and file changes to approve come to it as Allow / Always allow / Deny (or, with Bypass, run without asking), its questions are answered there, and it can be stopped and resumed. Needs the box upgraded.
-
Labs: a structured Claude agent's conversation can be resumed. When one ends, its pane offers Resume the conversation rather than a fresh start, and once its session is closed a worktree's menu lists its conversations under Resume a conversation: each carries on where it left off, earlier turns and permission mode included. Needs the box upgraded.
-
Labs: a structured Claude agent opens in Claude Code's own terminal and comes back. The Terminal switch on a structured agent's pane stops it and opens the same conversation in Claude Code's screen, for
/login,/memory, its menus and anything else only it does; Conversation brings it back to the chat, with what was said in the terminal in its place. Needs the box upgraded. -
Labs: structured Claude agents. Turn on Structured agents in Settings › Labs and new Claude Code agents run without their terminal: the box drives Claude Code through its stream and keeps its own record of the conversation, so replies stream in as they are written, nothing is read off a screen, and an agent keeps working while your laptop is away or the box's berthd restarts. Permissions come to the chat as Allow / Always allow / Deny, or with Permissions: Bypass (the default) the box allows them for you; questions and plans always wait for you. They open as conversations. Needs the box upgraded.
-
Doctor no longer says berthd runs "not as a service" when it does.
berthd installputs your login shell's PATH in the service, and doctor compared the service with one that had no PATH, so on a Mac it always asked you to runberthd installagain. Now any PATH counts; the program, its arguments and the rest must still match.
v0.3.16
-
Typing
/in a chat lists your skills. A bare/used to stop after the agent's first 60 commands, which hid every skill and plugin command (and the agent's last few) until you typed part of a name. It now lists everything: your own commands and skills first, then the agent's, then plugins'. And if the list couldn't be read when the chat opened, Shipyard asks the box again instead of leaving the menu closed. -
An agent's browser opens beside it. When an agent opens its browser in the worktree you're looking at, its live view now opens to the right of the agent, so you see the chat and the page side by side without clicking Agent's view · live. If the tab already has a browser pane, that pane switches to the agent's view instead. Your keyboard stays with the agent, and nothing opens while you're looking at another worktree. Turn it off in Settings → General.
v0.3.15
Both need each box's new berthd, and the first the Database plugin's
update (Settings → Plugins).
- Database finds a worktree's database in its
.env. Until now the Database tab only showed a database once Shipyard had started one of the worktree's services with its address, so a project that keepsDATABASE_URLin.envand starts its own dev server saw "No database yet" for every worktree. Now, when you open the tab and no service has said, the box reads the worktree's.env,.env.development,.env.localand.env.development.localand takes theDATABASE_URLorPOSTGRES_URLit finds, if it points at a Postgres on the box itself. The tab says where a database came from, the refresh button looks in the files again, and when there is none it says which files it read and why a variable it found isn't shown. And when nothing says which database a worktree has, Connect a database… takes its address: it has to be a Postgres on the box itself, it is kept on the box with the others, and Forget… removes it. See which databases appear. - Leadline records queries for projects that keep
DATABASE_URLin.env. When Shipyard starts a worktree's service and its own environment for it has noDATABASE_URLorPOSTGRES_URL, it now takes the one in the worktree's.envfiles and starts the service with Leadline's proxy in front of it, so Queries fills in without moving the address into Shipyard's config. A variable Shipyard's environment already has still wins, and the files are never changed. An app that overrides its environment from.envconnects round Leadline. See which databases Leadline sees.
v0.3.14
Database, Leadline and worktree new --no-setup need each box's new
berthd: use Update all in the status bar, or berth upgrade BOX.
- Database: a worktree's own development database, as a tab. A new
plugin in Settings → Plugins → Browse adds a Database tab to a
worktree: its tables with their rows in a grid you can sort, filter and
page through, each table's structure, and a console for SQL. It opens
read-only. Unlock writes asks once and lasts 15 minutes; then you can
edit cells, add and delete rows, review every change as old → new and
apply them together, or run statements that write. It is off until you
turn it on for a box, which installs a small program there
(
berth-database) after asking, and it only shows a Postgres on the box itself that the worktree's services were started with. See Database; plugin authors get aSqlEditorand the per-box switch for a plugin's program,HelperBox, in the plugin SDK. - A plugin's program stops when the plugin does. Turning off a plugin that runs a program on your boxes, such as Leadline, now stops that program and removes it from every box it runs on, keeping what it recorded; Remove… removes that too. Shipyard asks first and names the boxes, and a box that is offline is done when it next connects. Before, the program kept running and recording, with nothing in the app to stop it. Turning the plugin on again starts nothing: you turn it on for each box again. An update that keeps the same program leaves it running while Shipyard asks you about the new version, with Stop it in the plugin's row. See a helper does not outlive its plugin.
- Removing a worktree tells Leadline. Leadline closes the connection it
kept in front of that worktree's database and deletes what it recorded
for it, however the worktree was removed; before, it kept both until its
data was removed. Plugin authors: a helper can name a
detachpath. - Turning Leadline off again still names the services to restart. Turned off, on and off again without restarting a service, the second time named none, though the service was sending to Leadline again.
- The sidebar says when a box's skills are missing or out of date. On a box where Shipyard's skills were never installed, or after a box upgrade leaves them older than its berthd ships, a quiet note at the foot of the sidebar names the box ("devl: Shipyard's skills aren't installed for Claude Code", "gpu: 3 skills are out of date") with two ways out. Install now installs or updates them all for the box's user and the note goes; if a hook refuses or the box drops off, it says why. Snooze puts that box's note away until tomorrow, for a week or for a month, or Don't remind me for that box, which Settings → Agents turns back on. No toast, no sound. See the sidebar's reminder.
- The website's header is now coss.com's: "coss.com shipyard", with Docs, Download, GitHub and its stars, and "Close the laptop" as an icon, and coss.com's lines down the sides of the page and across the top of each chapter.
- Review no longer keeps a box busy while an item is open. With an item open, Review read the box's inbox and the item's diff again every couple of seconds for as long as it showed, each read setting off the next. It now reads them once, and again only when something changes.
- Make a worktree without waiting for its setup.
worktree new --no-setupcreates it at once, without the setup script or autostart services, andworktree setup LOC/NAME [--wait]runs them later. - Leadline: what your code did in development, as tabs on a worktree. A
new plugin in Settings → Plugins → Browse adds Queries,
Traces, Logs and Endpoints to a worktree: every database
query grouped by shape with its N+1s flagged, each request as a
waterfall with its SQL and logs, and per-route latency and queries per
request. Send to the agent writes a slow query, an N+1 or a failing
endpoint up for the worktree's agent, and shows you the text before it
goes. It is off until you turn it on for a box, which installs a small
program there (
berth-leadline) after asking; what it records stays on that box, for 72 hours. Queries to a Postgres on the box are recorded with no changes to your app; traces and logs need an OpenTelemetry SDK in it. A plugin that carries a program like this says so, by the program's name, before you allow it and in its row in Settings → Plugins. See Leadline; plugin authors getberth.helperand aChartcomponent in the plugin SDK. - Plugins can be installed from the app. Settings → Plugins has a new Browse list of plugins that aren't built in, starting with Hello ports. Install puts one under Installed, off: it runs only once you allow it, like any plugin of your own, and after an update Shipyard asks again. Remove deletes one you installed this way. To list a plugin of yours, see publish it to the listing.
- Shipyard has its own mark. The B from when the app was called Berth is gone: the icon is now a hull seen bow-on, resting on its keel blocks, with the amber lamp lit inside. It is on the app and its dock icon, the phone app, the site and the docs.
- Linear issues in Shipyard. Paste a personal API key in Settings → Integrations, and Issues gets a Linear side beside GitHub: what is assigned to you or a team's open issues, each with its description and comments, and Start an agent on any of them, in the project you choose, on the branch Linear suggests for the issue. A row then shows how its agent is doing. Home has a Linear widget for what is assigned to you. The key stays in a file only you can read on this computer: the laptop agent makes every call to Linear, and the app, plugins and boxes never see it. See Linear.
worktree newcan check out a pull request. Pass--pr Nwith the pull request's head branch as--branch, and the worktree starts on that branch as origin has it. A pull request from a fork is fetched frompull/N/head, or from the ref you give with--ref.- Links in a terminal open with ⌘-click, and say so. A link used to underline and show the pointing hand on a plain hover, where a click did nothing, and links a program made with its own text (OSC 8) never opened at all. Now resting on a link shows "⌘-click to open", the hand appears while ⌘ is held, and ⌘-click opens the address in your browser, OSC 8 links included (Ctrl-click in the Linux app). A long link that wraps over several rows, such as Codex's ChatGPT sign-in link, opens whole from any of its rows; before, it couldn't be clicked at all.
- A home widget's ⋯ menu opens again. It broke the home screen with "Base UI error #31" instead of showing its Size and Refresh options.
- A Browser tab on a port with nothing listening no longer crashes the
app. Opening
http://localhost:3000before its dev server started closed Shipyard half a second later, and again on every launch after, since the tab came back. - The laptop agent is lighter per request. Proxied pages, port forwards and calls to your boxes allocate a fraction of what they did, and the kits and team suggestions ask all your boxes at once.
- Shipyard opens faster. It now loads four files before it first paints instead of about three hundred, and settings, dialogs and the browser pane load just after start. In our tests, the window drew in about a fifth of the time.
- Chats keep up while an agent writes. A streamed reply costs about a third of the work it did on the page, and typing in a chat, Home's composer or the sidebar's menus stays quick with hundreds of worktrees.
- The app asks the agent and your boxes less. An idle chat makes about half the requests it did, starting reads every box once, and a box's large answers cross the link to your laptop about ten times smaller.
- berthd uses less of a box. Following a chat, a reply as it streams and the agents' screens runs fewer tmux commands and less CPU; a chat's own poll takes about a quarter of what it did. Opening a long conversation uses a tenth of the memory, and a cache that grew with every conversation now stays small.
- The Mac app is 20% smaller.
v0.3.13
Most of this needs each box's new berthd: use Update all in the status
bar, or berth upgrade BOX.
-
The Review in Shipyard button is for UI work. berthd adds it only to a PR whose files change something a browser shows, agents add it only for such changes, and the button is the words alone, without a mark.
-
Shipyard notices when your team has a setup. If you added a project from
acmeyourself and acme publishes a team setup, Shipyard says so quietly: a small icon on the project's row and a line in its menu, a line in Project settings saying what it sets up for the repo, a card on Home once, and Team setup for acme in ⌘K. Review… opens the setup with your project used as it is; nothing runs until you choose. Not now and Don't suggest again for acme put it away. The laptop checks only your projects' owners, with your owngh, at most daily (weekly for a person's account), and never offline. The project menu also has Copy team setup link. See how Shipyard notices your team's setup. -
Just the kit from a team setup. Use acme's kit for this project… applies only a project's kit from
acme/.berthto the project you have: no box steps, sudo, clone or 1Password. The sheet lists the tools the kit needs and whether the box has them; the shared keys and the first-time setup are each an opt-in. The kit then follows the team's newer commits as an update to review. See just the kit. -
Team setup uses the clones you already have. Before cloning a repo, Shipyard looks for one on the box: a project you added already, or a clone in
~/code,~/work,~/srcand the like (matched by its origin, reading nothing else). The page shows "Found ~/work/acme-web … · Use this / Clone a fresh copy", and using it changes nothing in it: no checkout, stash, reset or pull, and your.envand uncommitted changes stay as they are. Its existing worktrees show in the sidebar and are set up the first time you open each.berth team setup … --use web=PATHdoes the same from the command line. See repositories you already cloned. -
Review links can open logged in, at a page. Add
&as=pro@acme.test&path=/event-typesto a review link (or--asand--pathtoberth review), and the review opens in its Browser tab logged in as that dev user, at that page, when the project's login lists them; otherwise it opens without logging in, and the sheet says why. Copy review link as… in the worktree menu makes one. See review a PR. -
A Review in Shipyard button on pull requests. GitHub drops
berth://links from PR descriptions, so a review link is now a button that goes throughhttps://berthd.app/review, a small page that checks the link and opens the review in Shipyard (with a link to the PR on GitHub, and to install Shipyard). Agents add it to the PRs they open, logged in as one of the project's dev users and at the page their change affects (berthd login userslists them). Turn on"review_button"for a project (its.berth/config.json, Project settings, or the team'steam.json) and berthd adds it to each new PR from the project's worktrees by itself, once, without touching the rest of the description; never to a fork's PR or one from outside the org. -
Review a PR in one click, on your own box. A review link,
berth://review?repo=acme/shop&pr=42, orberth review acme/shop#42, opens a review sheet: the PR, its author, the exact commit, the box, what will run there, which keys it gets, and what in the PR changes setup. Review makes a worktree at that commit, set up by your team's kit, and opens it with the dev server in a Browser tab and the Diff beside it. Only your team's projects, and only PRs by members and collaborators from the repository itself; never a fork's. The PR's own.berth/never runs, the review never gets your own keys, and it cleans itself up when the PR is merged or closed, or after a week unused. Copy review link is in every worktree's menu whose branch has a PR. See review a PR. -
Open a worktree logged in as a dev user. A kit, or a project's config, can name a login script and the users it logs in (
"login": {"script", "users", "any"}). The Browser tab's Log in as… menu and ⌘K list them; picking one runs the script against that worktree's own dev server and database and sets its session for that worktree's address alone. Agents get the same withberthd browser open --as EMAIL, andberthd shots compare --as EMAILshoots both sides logged in. A link works too:http://WT.LOC.BOX.localhost:1377/__berth/login?as=EMAIL&next=/path. See kits.
v0.3.12
-
The Linux alpha's downloads. v0.3.11's Linux build stopped before it began (Ubuntu 22.04, which it builds on for older systems, has no
minisign), so v0.3.11 shipped for the Mac only. This release carries the AppImage and.debfor x86-64 and adds Linux to the update feed; see Linux (alpha). -
Install with Homebrew.
brew tap cosscom/shipyard https://github.com/cosscom/shipyard, thenbrew install --cask cosscom/shipyard/shipyard, installs the Mac app and putsberthon yourPATH. -
Handed-off worktrees nest under the one they came from. When an agent starts a task, or hands its work off, the new worktree shows under the agent's own in the sidebar, folded behind a "1 child" pill.
berthd worktree new … --parent NAMEnests one by hand, such as a stacked branch under the branch it builds on. -
GitHub flows see every worktree's PR. One poll looks at no more than 20 PRs, and it always started from the same worktree, so on a box with more worktrees than that the rest were never looked at: a merge, review or failed check there started nothing. Each poll now starts where the last one stopped. Merged and closed PRs no longer take a turn every poll either; a flow looks at them again once an hour, in case one is reopened. GitHub triggers.
v0.3.11
Most of this needs each box's new berthd: use Update all in the status
bar, or berth upgrade BOX. Agents keep running through the update.
- More than one way to each box. Shipyard can now reach a box over SSH
and at other addresses as well as over Tailscale, and sends new requests
over the fastest route that works. When the route in use stops answering,
requests move to another within seconds; open terminals follow once it is
clearly down, without closing, and keys typed meanwhile still arrive. A
box you added over SSH gets its SSH route on its own. Settings › Boxes
shows each box's routes with their latency ("via SSH, 24 ms · Tailscale,
140 ms"), and turns them on or off;
berth boxesshows the route in use andberth doctor BOXlists them all. - Sessions clean up after themselves. Ending a session now stops everything it started, such as a repository's Playwright tests and their browsers or a dev server, even ones left running in the background after the agent was killed. On Linux with systemd each new session runs in a scope of its own, which also lets it outlive berthd as before; elsewhere Shipyard finds the session's processes by their tree (and, on Linux, the session's marker). Your own processes and tmux are never touched.
- See and stop a box's browsers. Click a box's memory meter in the
status bar to see every Chromium, Chrome and Firefox on it, with who
started it (Shipyard's agent browser, an agent-browser session, a
session's tests, one left behind, or yours), its CPU, memory and age, and
Stop for Shipyard's and its sessions'. The sessions using the most
memory are below. Settings → Boxes shows the same, and agents see it with
berthd psand inberthd doctor. - A memory limit per session. Settings → Boxes can hold each session on a box under a limit, such as 12 GB (Linux with systemd). A session near it is slowed down, never stopped, and its chat and tab say "using 11.8 GB, near its 12 GB limit".
- A lighter agent browser. It closes after 5 idle minutes instead of 10
(agent-browser sessions too), keeps to two renderer processes, and caps a
page's JavaScript heap at 512 MB, which saves about 35 MB per browser
(
BERTH_BROWSER_JS_HEAP_MBraises it). - A slow or relayed link to a box stays connected and says why. A box
reached through a Tailscale relay far away (80 ms, then 2 s) used to flip
to offline and back: the agent called a box away after one health check
with no answer, and checked it whenever a request had waited 5 s. Now a
slow or unanswered check marks the box slow, not offline: requests to
it still go through and its connection, with the terminals on it, is
kept. Only a second failure in a row takes it away (a refused connection,
a changed key or berthd stopping still does at once), so a box that
really went shows Reconnecting in about 20 s. A request now waits 10 s
before it has the box checked. The status bar and sidebar say "slow",
quietly; Settings › Boxes shows the latency, why it's slow, and when
Tailscale relays the box: "Relayed through Tailscale's New York server:
no direct connection", with what fixes it.
berth doctor BOXsays so too, and the agent's log has a line for every change of a box's state, with why. A relayed link. berth pingsays a box's latency. It printed one number that was connecting, TLS and a request together (about three round trips), so a box 100 ms away read as 300 ms. It now says how long connecting took, then pings on the open connection, as the app uses it:cal: 88ms 90ms 118ms (min 88ms, avg 99ms, max 118ms on the open connection).-c Nsets how many.- Typing in a terminal keeps up on a slow link. On a box 80 ms to 2 s away (a relayed Tailscale link), a shell showed each key only once the box had echoed it. Now what you type shows at once, underlined until the box's echo agrees, as in Mosh; anything the box draws differently takes every prediction back, and password prompts never show one. It turns on by itself when the round trip is over about 60 ms: Settings → Terminal → Predict typing has Always and Never too. See Terminals.
- The Browser tab's address follows the page. An app that changes pages
on the client (Next.js, React Router:
history.pushState) left the address bar on the first URL, since nothing loaded. It now shows where the page is, within half a second. - Every agent account gets Shipyard's hooks and skills. A Claude Code or
Codex session started on another account (the Usage plugin's accounts,
which set
CLAUDE_CONFIG_DIRorCODEX_HOME) ran without Shipyard's hooks, so the box guessed from the screen when it was working, done or needed you, and without Shipyard's skills.berthd integrations installnow puts them in every account folder (the default, those in~/.berth/accounts, and any the box or a project picks), Add account… puts them in the new one, and berthd adds them to an account a session starts on before its agent runs.berthd integrationsandberthd doctorlist them per account. See Agent integrations. - A chat on another account reads its conversation. An agent started on
another Claude Code or Codex account (the Usage plugin's accounts, which
set
CLAUDE_CONFIG_DIRorCODEX_HOME) wrote its conversation to that account's folder, and the box only looked in its own: the chat said "Its conversation can't be read here yet" and emptied after every message. The box now reads each session's conversation from the account it started on. - Berth is now Shipyard. A new name, nothing else: the same app, the
same CLI and the same boxes, with your data where it was. The commands are
still
berthandberthd, and their files (~/.berth, a repository's.berth/,BERTH_*variables,berth://links) keep their names for now. A new download isShipyard.app; an installedBerth.appupdates in place and keeps its folder name. Releases before this one went out as Berth. - Shipyard for Linux, an alpha. The desktop app now ships for x86-64 Linux as an AppImage, which updates itself from the same feed as the Mac app, and as a .deb. It's an alpha: less tested than the Mac app, and its sidebar and Settings → About say so, with Report a Linux bug in About. Ctrl stands in for ⌘ (in a terminal, add Shift: Ctrl+Shift+T, and Ctrl+Shift+C and V copy and paste); there's no menu bar; Start at login makes the agent a systemd user service. arm64 isn't built yet. Linux (alpha).
v0.3.10
-
What's new, once after an update. The Mac app updates itself quietly, so its new things went unseen. After an update, a note at the foot of the sidebar says what's new (with the sidebar folded, the card opens instead). The card shows the release's highlights one at a time, each with a small picture of it in the app and Show me, which goes there: the worktree's artifacts board, a visual diff, a Browser tab with its console open, or renaming a worktree; the rest show their keys or their guide. ← and → page through it, Esc closes it. A new install doesn't get it, and × on the note puts it away for good. Settings → About → What's new and ⌘K open it again. Updates.
-
A team setup's repos are ready when it says so. Each repo's first-time setup (its init) runs in a terminal of its own on the box, which the Team setup page never showed. An init that stopped at a question, such as corepack's "Do you want to continue? [Y/n]" before it downloads yarn, left the repo setting up until opening its main checkout happened to show that terminal; answering there finished the setup, which is why worktrees only worked after Open main checkout. Now corepack doesn't ask (in inits and worktree setup scripts). A repo's row shows its init's last line and a Terminal button. When the init asks anything, the row says waiting for you, the status card names the repo, and Answer in its terminal opens it. Open main checkout is a way to open the repo again, not a step. Progress, failure and Retry.
-
An agent CLI you installed with npm is found. Claude Code installed with
npm i -g @anthropic-ai/claude-codeunder nvm (or fnm, volta, bun, pnpm, a custom npm prefix or Homebrew) wasn't picked up on a new box, where the one Add agents… installs was: berthd runs as a service, and only your interactive shell's~/.bashrcknows where nvm put it. berthd now asks your shell as a terminal starts it ($SHELL -lic, reading only its answer, past any banner, and giving up after ten seconds), then looks in the folders those installers use. With both installs, the one your shell finds first is used. Agents start with it and thePATHit was found with, so itsnodeis found too.berthd doctorand Settings → Boxes say which one, such asClaude Code 2.1.3 · ~/.nvm/…/bin/claude (npm), with Look again for one installed since;berthd agents installno longer installs a second copy, andberthd integrations installfinds it the same way. Agent CLIs you installed yourself. -
The agent's browser has the size the agent chooses. It opens at 1920×1080 now (it was 1280×800), and an agent sets another with
berthd browser resize 390x844 --scale 3, a preset (phone,phone-maxandtabletas in Preview,laptop,desktop), oropen --size … --scale …. The size is kept for the worktree until it changes, a running browser changes at once, and widths from 320 to 3840, heights from 240 to 2160 and scales from 1 to 3 are allowed, with an error that says the range otherwise.open,status,list,doctorandGET /v1/browser/healthsay the size;shotfollows it (--nativefor the page's own pixels). In the app, Agent's view says the size and the zoom it shows the page at, and fits the page to the pane without enlarging it past its own size, so a 2x page stays sharp. Theberth-browserskill tells agents to check a phone size too. Agent browser -
Files an agent's browser downloads land in the worktree's
.berth/browser/, as the guide says, rather than the folder berthd started in. -
A slow or dropping link to a box says so, and recovers by itself. Tested against boxes 200–900 ms away, links with jitter, a bandwidth cap, stalls, resets, outages of 10 s, 60 s and 5 minutes, and a laptop asleep. A box that drops off now reads Reconnecting to devl… Next try in 6s in its terminals, chats and a team setup in progress, with Try now (which checks the box at once; it used to only re-read what the agent already knew). The agent tries an away box after about 2 s, then 4 and every 8 s, each with jitter (never less often than before), checks a box at once when a request to it has had no answer for 5 s, and answers requests for an away box at once rather than after a 15 s dial, so ⌘P, Files and the rest stop spinning. Reads now give up after 30 s with "devl isn't answering" instead of spinning on, and ⌘P says why in words and searches again when the box is back. Typing into a terminal that gets nothing back for 4 s shows Waiting for devl to answer…. A chat message sent again after a dropped answer carries the same idempotency key, so it is typed once, not twice. The app's event stream to the agent reconnects with jittered backoff and is dropped and opened again after 45 s of silence (a laptop that slept), and refreshes that land together after a reconnect run as one instead of piling up. Slow and dropping networks.
-
Berth works from the keyboard and with a screen reader. Under a chat, its agent's terminal no longer takes Tab and keeps it. Closing ⌘K, ⌘P or a rename gives the keyboard back instead of dropping it; a file opened from ⌘P takes the keyboard, as do the Files panel (⌘⇧E) and the console drawer (⌘⌥I), whose tabs switch with the arrows and whose edge sizes with ↑ and ↓. Shift-F10 opens a tab's menu, ← and → fold a project in the sidebar, answering a permission or a question keeps the keyboard, and focus rings show in every theme (3:1, and muted text 4.5:1 on menus and highlighted rows: 22 themes moved a shade; code in the File tab and in chats reads at 4.5:1 too, each colour a shade darker or lighter). Screen readers hear who said each message in a chat, and hear an agent that finishes, needs you or fails, alerts first. ⌘K now lists every shortcut's action, each part of Settings, Review, a Preview tab, Rename this worktree and Do not disturb; the shortcuts sheet adds F2, ⌘F, ↵, esc, ← → and ⇧F10. With Reduce motion on in macOS, spinners, shimmer and pulses hold still and nothing slides.
-
Artifacts: charts, tables, diagrams and small pages from your agents, live. An agent writes a
berth.chartJSON, a CSV, Mermaid, Markdown or one HTML file and runsberthd artifact add; its chat shows a card with a live thumbnail and the headline its data gives ("/search 1,240 → 410 ms (−67%)", "3 failed of 16"). Open it as a tab, or ⌘-click to open it beside the chat; the tab has a Live pill that pulses when the agent rewrites the file, its versions (the first and the newest eight) and its source. N artifacts under the reply box, or the grid button in the worktree's toolbar, opens the worktree's board, filtered by Charts, Tables, Diagrams, Pages and Notes. Charts are drawn with bklit UI in your theme, in all 23; a page runs on its own origin with no network and no access to Berth. The box keeps them outside the worktree, checks each version, and refuses one that looks like a secret. Theberth-artifactsskill teaches Claude Code and Codex when and how (berthd integrations install). The claude.ai chip now reads N published. Artifacts. -
Visual diffs: what a change did to your pages. After a UI change, an agent runs
berthd shots compare: the worktree's pages and main's are shot at 375, 768 and 1280 pixels wide in a short-lived Chromium of their own (reduced motion, animations settled, fonts and images loaded, lazy content scrolled in, UTC), diffed pixel by pixel with content that only moved counted as moved, and kept as an artifact in its chat, in about 5 seconds for 6 pages × 3 sizes. The tab opens on the biggest change: a before/after Slider, Side by side, Flicker or Onion skin, a Heatmap with Glow, numbered regions named by their element, Changes ‹ n of N › (n and p), All shots in a grid, and chips for a page that now scrolls sideways, fails or is new. Accept as baseline keeps the after-shots; comparing again with--base acceptedgives the all clear. Masks and defaults go in.berth/config.json'sshots;--color-scheme bothshoots light and dark; with main's dev server stopped it compares with the turn-start baseline (berthd shots baseline) and says so. The agent reads about 450 tokens. Theberth-visual-diffskill teaches Claude Code and Codex to check before saying done. Beside the chat, in a narrow pane, its tab's header folds to one line (the title, a version menu, Source and Board) so the before/after starts near the top. Visual diffs. -
Berth rests when nothing happens. Left open with a couple of chats, two Browser tabs and their drawers, a Preview tab and the Files panel, the app asked the agent about 170 times a minute and kept its main thread busy about 12% of the time, hidden or not, because every terminal drew on every frame whether or not it showed. Now a terminal in a hidden tab draws nothing and gets its output a second at a time (the agent batches it), and one on screen draws only when something changes: with a dozen terminals printing behind other tabs, the main thread went from most of a core to about a tenth of one, typing in the one in front echoes in about 30 ms instead of 400, and switching to one takes 30 ms instead of 400. At rest the app asks about 55 times a minute, and nothing while the window is hidden or the Mac sleeps; it catches up the moment you come back. Spinners and shimmers hold still once the window has been in the background for 30 seconds. Closing a terminal now frees its memory (it kept all of it before), a full Console drawer no longer redraws every row for each new line, xterm.js (the fallback renderer) no longer stalls with noisy terminals in background tabs, and an idle box's daemon no longer wakes thirty times a second.
pnpm perf:soakandpnpm perf:terminalsmeasure it; the smoke suite checks it stays so. -
Give a worktree a name you can read. A worktree made from a pasted link is called
https-linear-app-acme. Rename… in its menu, double-click or F2 on its sidebar row (Enter saves, Esc cancels), or right-click or F2 on its tile in the folded rail gives it a display name; the field says the branch stays as it is, and offers the name its agent's task gives. It is a label only: the branch and folder don't move, and an empty name shows the worktree's own again. The box keeps it, so your other laptops, the phone andberth worktree rename BOX/LOC/NAME "Name"see it too. It shows in the sidebar, the rail, tab groups, ⌘K and ⌘P, Home, Review, notifications, the status bar (new: the worktree in front, click to rename) and the window's title, with the worktree's own name in its tooltip and beside it when the sidebar is wide; ⌘K finds either. A box that isn't updated keeps it on this laptop only, and says so once. Display names. -
Drag the sidebar wider or narrower. Its right edge sizes it from 200px, where a worktree's name still reads, to 420px (and never more than 40% of the window), and the width is kept. Double-click the edge for the default 240px; focus it for ← and → (Shift for bigger steps), Home and End. Dragged past its narrowest it folds to the rail, as in VS Code, and dragging the rail's edge out opens it again; Esc mid-drag puts it back. Terminals and the Browser tab follow as you drag.
-
Long chats stay quick. In a chat of 2,000 turns, each update of a reply being written took about 57 ms, as every row was drawn again and every message re-read for ⌘F; now only the reply's own row draws, in about 15 ms, the same as in a short chat. Typing in the reply box no longer makes the whole window restyle on each key (about 8 ms a key, at any length). Scrolling a long chat from top to bottom quickly no longer drops half its frames, and a chat holds half the memory. Words you select stay selected while the chat scrolls past them, so dragging or Shift-clicking across many messages copies all of them, and screen readers hear the chat as a list that says where each message is. Conversation view.
-
The sidebar stays quick with hundreds of worktrees. With 300 worktrees, each with an agent, every row used to build its right-click menu, its ⋯ and + menus and its tooltips up front: about 1.2 s of work as the app started, and 120 MB of memory. Now the sidebar and the folded rail share one menu and one tooltip, and a row draws its buttons when you first point at it or tab to it: the start costs no more than with a handful of worktrees, memory is about 20 MB, and folding a project, renaming a row or going to Home no longer draws every row again. Agents' spinning and pulsing state icons rest while scrolled out of sight, which kept the app busy even when nothing happened, and dragging the sidebar's edge no longer restyles the whole window on every frame (it stuttered at a few frames a second). Right-click, Shift+F10, the menu key, F2 and double-click work as before.
-
The Browser tab stops spinning once a page has loaded. A page that adds a frame after it loads (a chat widget, an embed) used to leave the spinner on for good. Only the page itself counts now, and a load that never starts gives up after ten seconds. A page whose own code keeps reloading it (four times in a row) says so, with Open in your browser. Next.js 16.3 reloads a page once in WebKit after a redirect to a page with a Cross-Origin-Opener-Policy header; Next 16.4 fixes it, or set
experimental.reactDebugChannel: false. -
Messages from other agents are theirs, not yours. Claude Code hands a helper's hand-back, a teammate's message, another session's, the lead's note and its own background notices to the model as if they were prompts. The chat showed some as your own bubble (tags, JSON and reminders included) and dropped the rest, so Claude answered reports you never saw. Now each is a card beside Berth's reports: the sender's avatar and colour, its name and kind, Report, Question, Update, Instruction or Failed, how long it took, "to Claude", and Open for a helper's tab. Read report unfolds a report in place; one longer than a screen opens in its helper's tab or a sheet. A teammate's or session's question needs you until it's answered, with Reply. Background pings are one line, successes close together fold into "✓ 3 finished", failures always show (×2 when repeated). A message you sent mid-turn is marked "Sent while Claude was working", and
<system-reminder>notes are gone. Teammates join the Crew strip, helpers that reported back get a Report badge, and Berth's own report cards take the same head. Needs the box updated; an older box shows these as before. Conversation view. -
The crew's times fit. A helper out for 10 minutes or more no longer wraps its time onto two lines ("12m / 00s").
-
Developer tools in the Browser tab. Inspect opens WebKit's Web Inspector for the page, in its own window, and the page's right-click menu has Inspect Element, in release builds too (a private WebKit setting the App Store refuses; Berth ships outside it). Console and network (⌘⌥I) opens a drawer under the page: its console since it loaded, uncaught errors and rejections with their stacks, and its requests from the laptop proxy's log, failed ones marked. Send to agent on an error or a failed request hands the worktree's agent the error, its stack or the request and its answer, and the page's address. The Browser tab and its toolbar count the page's errors. Watching the agent's browser, the drawer shows its console and failed requests. Browser tab
v0.3.9
-
A helper opens in its own tab. Click a helper in the crew, or in "Sent out N helpers", and its whole conversation opens as a tab just after its chat's: its name, a spinner while it works and a check once it's back, and the agent that sent it out a click away. Click again and that tab comes to the front. ⌘-click opens it beside the chat in a split, and ⌥-click (or the row's eye) still peeks in the sheet, which now has Open in tab. The tab splits, drags and groups like any other, and is there again after a restart; a helper whose record is gone, or whose box is offline, says so.
-
agent-browser no longer piles up on a box. Agents that drive Vercel's agent-browser CLI instead of Berth's browser left a daemon and a headless Chrome running for every session they forgot to close, and on a busy box those Chromes ate the CPU. Every session Berth starts now gets
AGENT_BROWSER_IDLE_TIMEOUT_MS=600000, so an idle one closes after 10 minutes (set it in the box'senv.jsonor the project'senvto change it,0to turn it off). When a Berth session stops, exits or its worktree is archived, Berth closes the agent-browser sessions it started, and only those: one run by hand is never touched.berthd doctorlists any left behind andberthd browser reapcloses them. Theberth-browserskill now tells agents to useberthd browserfor the worktree's pages, and to runagent-browser closewhen they use agent-browser anyway. Agent browser. -
The crew sits in its conversation. The helpers an agent sends out now dock above its reply box, beside the task list, in each chat pane for its own agent. Before, the card sat in the window's corner, and in a narrow pane it hung over the reply box and the Stop button. Folded it is one line (how many, how many are working, what the first is doing); it folds itself a few seconds after every helper is back. Loops keep the corner.
-
Adding a box is quiet. Add a box over SSH, or Install and pair on a tailnet machine, opens a compact dialog with a short checklist, and the agents to install sit inline beside the host (Claude Code ticked, remembered for the next box). Berth installs berthd, its own tmux, the agents and their integrations by itself: no plan to read and no Enter to press. A terminal opens in the dialog only for a step that truly needs
sudo's password (git missing, or keeping berthd running after you log out where the box allows that only as root), and only for that step; lingering is tried without sudo first, and when it alone needs the password the dialog asks, with Skip.berth add sshis quiet the same way;--guidedshows the whole plan and runs it in one terminal, as adding a box from Team setup does (the guided install, below). -
1Password is optional in Team setup. When a team's shared keys are 1Password references, the Keys check offers Use 1Password or Skip 1Password. Skipped, each shared key is typed once (or left blank), no
op://reference reaches the box andopis never called for them; blank keys are listed as missing on You're set up and in the project's settings, and Use 1Password turns it on later. A team can require it with"onepassword": "required"inteam.json(the default is"optional").berth team setup --no-1passworddoes the same from the command line. -
The guided install. Adding a box from Team setup first shows the plan, full screen: what will run, step by step, the exact commands a click away, which steps may need
sudo, the team's own steps after Berth's, and an Agents choice (Claude Code ticked; Codex, Cursor Agent and OpenCode to add; remembered for the next box). Thenberth add ssh --guidedruns in a terminal here, beside a checklist: press Enter to start, and type your password whensudoasks for it on the box (it goes to the box and nowhere else; Berth never sees or keeps it). A failed step shows its command as a copyable line and Retry from here; the steps before it are kept. When the box is ready, the team's setup starts on it in the same screen, and Back to Team setup goes on from there.berth add ssh --guideddoes the same in your own terminal: the plan, then Enter;--yesand scripts stop with the command to run instead, never a command made into a sentence. -
Berth brings its own tmux.
berthand the app carry a static tmux for Linux (amd64 and arm64, 1.3 MB, built from pinned sources byscripts/build-tmux.sh), uploaded to~/.local/binon a box without one, so tmux needs no sudo. git still comes from the package manager, in the guided terminal. berthd's service now has~/.local/binon itsPATHon Linux too. -
Berth installs the agent CLIs.
berthd agents install claude codexputs them in~/.local/binwithout sudo: Claude Code with its native installer, Codex from its pinned GitHub release with its sha256 checked, Cursor Agent and OpenCode with their own installers; each is skipped when it is there. Adding a box runs it, then installs their hooks and skills (berthd integrations install present), so there is nothing left to run by hand. Settings › Boxes › … › Add agents… adds more later, in the same view. Signing in stays yours. A team setup can name them too:"agents": ["claude", "codex"]inteam.jsonbecomes Berth's own step. -
Adding a box refuses to install as root, and asks before berthd listens on every interface on a box without a tailnet address.
-
Team setup. A GitHub org publishes how its engineers' boxes are set up in a repository called
.berth(team.json, schemaberth.team/v1, plus the box script and any kits), and a new engineer runs it from Berth: Joining a team? on the welcome screen, Set this box up for a team in Add a box, or a link,berth://team?org=<org>. One page reads like the org's repository (who published it, the commit, who changed it last), lists what you'll get with every exact command, and turns into the progress when you press Set up. GitHub goes through the GitHub CLI: your laptop'sghreads the setup and checks you can read each repo (no membership check; "You can't read acme/website: ask an admin · Set up the other 2"), and the box signs in with its owngh auth login, so your laptop's token never leaves it. Box steps run in a terminal on the box, as you; steps that need root ask for your password there, and Berth never sees it. Progress survives berthd restarting, a failed step keeps the ones before it and retries from itself, and each repo is set up from its own.berth/config.json(trusted as you reviewed it) or the team's kit, pinned to a commit. Shared keys are 1Password references, your own are asked for once and stay on your box. A newer commit of.berthis an update to review, with a diff, never run by itself. An org without a.berthgets a picker of its repos. A team setup also loads from a link to any repo, branch and folder (github.com/<owner>/<repo>/tree/<ref>/<folder>, orberth://team?src=…), to try one before the org publishes it; the page then names the repo and branch it came from.berth team show|setup|status|retrydoes the same from the command line. Acme's is the worked example, inexamples/team/acme-dot-berth. See team setup. -
Team setup signs 1Password in. A team setup whose shared keys are
op://references gets a Berth step of its own after GitHub's, 1Password on the box:op signinruns in the box's terminal, the page shows it inline while op asks (its questions count as waiting for you, as sudo's do), and berthd keeps op's session for its reads.berthd secret signindoes the same by hand, and--checkonly says whether the box can read 1Password. -
op never asks in a service's terminal. Berth runs op without a terminal, so a box that isn't signed in no longer leaves a service sitting at op's "add an account? [Y/n]": the read fails at once with "1Password isn't signed in on devbox: sign in from Team setup or run
berthd secret signinon the box", and the service starts without the key. -
tmux and git on a fresh box.
berthd install(which the install command andberth add sshboth run) installs tmux, and git if missing, when that needs no password: as root, with sudo that needs none, or with Homebrew. The install command offers to run it with sudo when it can ask; otherwise it stops before installing anything, with the command to run. The Team setup page says when the box has no tmux, with the command, and waits for it. -
box.settingsinteam.json. Values the team keeps in one place (versions, ports) reach the box script, and each project's init, asBERTH_SETTING_<NAME>. Strings only, checked;$fields are still comments. -
New groups reach what berthd starts. A step that adds you to a group (docker) used to leave berthd, and every terminal and service it started, without it until a reboot. The team's terminal now carries on with the new group, and berthd starts terminals, services, scripts and hooks through
sgwhile it lacks one;berthd doctorsays so. -
Kit requirements are checked as a login shell finds them. A kit's
requiresused berthd's own PATH, so a project warned "node is not installed" right after the box steps installed it with fnm. They are looked for through your login shell now, and checked again when a team setup's steps end. -
A project's init gets the kit. A team setup's
initscript getsBERTH_KIT_DIR,BERTH_LOCATION,BERTH_ROOT_PATHandBERTH_BOX, as kit scripts do, so it can reuse the kit's code. -
Team project ids are URL-safe. A project's id names it on the box and is part of every worktree's URL, so
team.jsonrefuses one with a dot (acme.com; usesite, withpathkeeping the folder). Repos picked from an org without a.berthget one (acme.com→acme-com). -
The example team setup is Acme's, a made-up org whose shop needs Node, Postgres and Redis: 8 box steps, 4 with sudo, versions in
box.settings, and an init that migrates and seeds.
v0.3.8
-
The Files panel. ⌘⇧E (or View → Files panel, or its button in the tab strip) shows the worktree's files at the right, beside whatever tab is in front, so you can watch what the agent touches while you read its Chat. All lists the worktree a folder at a time; Changed folds it to this turn's files, with
+/−lines and A/M/D letters. It opens on Changed while the agent works and on All otherwise, until you pick. Blue dots mark folders with changes inside, and the file the agent is writing gets a live mark while it works. Click a file to open it in a File tab; New file makes an empty one in the folder you are on, and is the only thing the panel does to files. Below a 1100px window it floats over the tabs and goes on esc. It stays open across restarts. The box lists a folder at a time (GET …/files?dir=), so the panel isn't limited to the 20,000 files ⌘P searches; past that, ⌘P also searches what the agent changed, and says so. See the Files panel. -
⌘P finds a file, and a File tab opens it. In any worktree, ⌘P lists the files the agent changed this turn first, with their
+/−lines, then the ones you opened lately; type to search every file, fuzzily, with letters in the file's name counting most. On a wide window the highlighted file shows beside the list at the agent's first change. ↵ opens a File tab, ⌥↵ opens it beside the pane, ⌘↵ in your editor. The File tab marks the agent's changes this turn in the margin (yours, unsaved, in grey), steps through them, saves with ⌘S, and is coloured by your theme's own syntax colours, as diffs are. When the agent writes the file under your unsaved edits, nothing is overwritten: a banner offers Reload, Keep mine or Compare (side by side in the diff view). Large and binary files say why they don't open; pictures show. See Go to file and File tabs. -
Prompts wait for an agent's trust question. In a folder it hasn't seen, Claude Code (and Codex) first asks whether to trust it, and that question takes keys only: a prompt typed into it was dropped, and the Enter after it answered the question, which with Claude Code's highlighted "No, exit" ended the session before its first prompt ran. The box now watches a new agent until it is at its prompt. A prompt sent while it asks is held and typed once you've answered and it has run the prompt it was started with (that one, given on its command line, was always kept); one sent before it has drawn waits a moment first. The chat shows the question in the agent's own screen, to answer there, with the first prompt marked as sent, and no longer offers a Yes or No that would type a word into it. Needs the box updated.
-
The chat streams while Claude writes. A reply now grows in the chat as it does in the terminal, read off Claude Code's screen, with a caret where the words end; when Claude Code records the message it takes the draft's place in the same spot, with nothing jumping or showing twice. A reply taller than the terminal keeps the start the chat already saw. Finished steps (tool calls, earlier words) also appear within a fraction of a second instead of every two seconds: the box says when an open chat's transcript changes. Codex chats show messages as they land, as before. Needs the box updated; older boxes keep working as they did. Settings › Appearance › Chat › "Show replies as they're written" turns the growing draft off.
-
An agent hears back when work it started is done. An agent that starts a task in a worktree, prompts another agent, or starts a run no longer has to keep polling to find out it finished. Berth remembers who asked and, when the work finishes, fails, needs a person or reaches a gate, types one short
<berth-notification>into the asking agent's session at its next idle: the agent's answer, the files it changed with their line counts, its branch and how to review and merge it, and where to read more. Several that land together arrive as one message; each is said once, never in a circle, never to an agent that's gone, and at most one every 15 seconds. The skills and MCP tools now tell agents to end their turn rather than wait, andnotify: false(or--no-notify) opts out. In the app the report reads as a card in the chat, with a button to open the agent it's about. See hearing back. -
A file dropped on the window no longer replaces the app. Dropping a file from Finder where nothing takes it used to open it in Berth's place, a whole-window markdown file with no way back. Now it is refused, and a hint says to drop it on a message box or a terminal. Home's composer, the ⌘N task box, Running agents, Fork and Send back now take dropped or pasted files, as the reply box does: they go up to the box and their paths go with the prompt (for a new worktree, into the project's main checkout until the worktree exists).
v0.3.7
-
The folded sidebar shows your agents, not initials. With the sidebar hidden (
⌘\), the rail lists agents by what they need: those that need you, then working, then done, then any on a box that went away, each lane headed by its glyph and count. Each agent is its own icon (Claude, Codex) in a ring of its state, over its worktree's name, edged in its tab group's colour when on screen. Hover or focus one for its task, what it waits for, its project, worktree and box; click to go to it. Lanes fold; arrow keys move between agents; nothing shows while nothing is happening. -
Ask about what you select in a chat. Select words in an agent's reply, a tool's output or a question, and a small bar offers Ask Claude about this (type and press Enter: it goes with the words quoted) or Quote in reply. Code blocks in replies now say their language and keep Copy in view, and waiting ("Reading the conversation…", the agent's working line) shows a small pixel grid with how long it has taken.
-
Home is yours: a grid of widgets under the composer. Needs you (with Allow once for a permission, right there), Working now with each agent's step, Recently finished, Recent areas, the Harbour, Pull requests and your review queue, CI failures on your worktrees' branches, Git activity, Agent usage, Boxes, Running services and Notes. Customize moves, resizes and removes them (by drag or keyboard), Add widget shows each one live with your data, and your layout is kept. Widgets read nothing while off screen. Plugins add their own with
berth.addHomeWidget. More -
Preview tabs: a page at every size at once. A worktree's dev server at Tailwind's breakpoints, phones and a tablet side by side, each at its real width, as a filmstrip, a grid or all fitted in view. Synced, opening a page or scrolling in one frame does it in all (clicks and typing too, if you like); isolated, each is its own. Force light or dark for all frames or one, rotate phones, add your own widths, and take one screenshot of every frame. Open one from + → Preview, the Run menu, Running in, or a Browser tab's toolbar. The laptop agent gives only Preview frames the script that does this; restart it after updating. More
-
A terminal on a box, from Home. ⌘T or New terminal with no worktree in focus asks which box (each with its state and load, the last one picked first; a box that's offline or needs a newer berthd says so), or goes straight to the only one online. The shell starts in the box user's home folder and opens as a tab over Home, after a devl · ~ chip; closing the last one gives you plain Home back. A box's menu in the sidebar has New terminal too. Boxes need this berthd (
session.home). -
Agents start on a Mac's own box again. Under launchd, tmux 3.7 had no locale and mangled the session list, so every agent on "this Mac" ended with "It's gone". tmux now always speaks UTF-8 there.
-
A new task opens in one tab, not two.
-
An update restarts the Berth agent. The agent outlives the app, so after an update it went on running the old version (and kept this Mac's box on the old berthd) until you restarted it or the Mac. Now the updated app restarts it as it starts, whether the app started it or it starts at login: it first finishes what it's doing, your agents keep running on their boxes, and the new agent brings this Mac's box up to date. A short note says Updated to vX.Y.Z.
berth agent restartdoes the same by hand. -
The chat no longer sticks on "Reading the conversation…". An agent at work sends an event with every tool call, and each one started the chat's read over, throwing away the answer on its way: on a box slower to answer than the agent's calls came, the chat never showed. Reads now carry on, and one the box never answers ends in a few seconds in Couldn't read the conversation with Retry.
-
Compare two worktrees (Labs). ⌘⌥C, Compare with… in a worktree's menu, or Compare with #N on an attempt in Review opens a Compare tab: both worktrees side by side in mirrored lanes, Chat, Diff, Preview and Terminal (⌥1–4), with each side's agent state and diff numbers, a hinge that swaps them (⌘⌥S), and Sync: the same path on both dev servers with each page's load time, the same file in both diffs, the same turn in both chats. A side that is archived or offline leaves the other to fill the tab, saying why. Tabs, group labels and panes are reachable from the keyboard and named with their worktree for screen readers, a colour is always shown with a name, and each theme sets worktree colours so they read on it. Labs.
-
Copy diagnostics. Settings → About → Help, and an error's Details, copy one short report to paste into a chat: the app's and the agent's versions,
berth doctor's checks, this Mac's box (its tools, tmux, agent CLIs andPATH), each box's version, capabilities and last error, the terminal renderer, theme, Labs, and the last errors and toasts with their codes. Tokens,op://references, emails and your username are taken out.berth doctor --reportprints the same text. -
A smoke suite for the app.
pnpm test:e2e(inapp/) runs the production build in Chromium on the mock fixtures, and CI runs it on every change;pnpm smoke:liveruns it against your own agent, read only.
v0.3.6
October 5, 2026.
Worktree pages open on a Mac's own box, in your browser and in Berth's.
- Worktree URLs work on a Mac's own box, and in the Browser tab. On
this Mac's box,
http://WORKTREE.LOCATION.BOX.localhost:1377/said "not found" for a dev server started in the worktree (npm starton 3000), since the box only placed a server by its port, and terminals never gotPORT. The box now reads each server's folder on macOS too, as Linux boxes always did, and a worktree's terminals and agents getPORT=$BERTH_PORTunless the project sets it. A URL with nothing running yet says so and how to start it, and a server started a moment ago opens at once. Separately, the release app's browser panes stayed blank on everyhttp://…localhostpage: macOS App Transport Security refused plain http in the bundled app (never inpnpm tauri dev); Berth.app now allows it for localhost names only. When terminals fall back to xterm.js, the app says why once, with Copy details, in Settings → Terminal, and inberth doctor. Boxes and the app need the update. Project config.
v0.3.5
October 5, 2026.
Agents start however long the prompt, and on a fresh Mac.
- Long prompts start. A task or session whose prompt was over about 16 KB failed with "tmux new-session: ." and its worktree was removed: the command went on tmux's command line, which tmux refuses. Each session's command now lives in a file only you can read, which its login shell runs, so a prompt can be up to a 2 MB request (the agent itself takes its first one as one argument: up to 128 KB on Linux, about 1 MB on a Mac, and a longer one is refused up front, saying so). A long message sent to a running agent goes through tmux's stdin, whole. Every tmux failure now says why (its exit status, or a timeout), never an empty reason. Boxes need an upgrade. Sessions.
- This Mac finds Homebrew's tmux. The Mac's own box runs under launchd,
whose bare PATH hid
/opt/homebrew/bin, so no agent could start there. berthd now looks for tmux in Homebrew's folders too, and its launchd plist carries your login shell's PATH (with Homebrew's and~/.local/bin); the agent rewrites an older plist on its next start.berthd doctorsays how to install tmux on that system. - Missing tmux or agent CLI gets a card, not a failure. Starting an
agent on a box without tmux or the agent's CLI shows how to install it:
onboarding and ⌘N show the command (
brew install tmuxon a Mac, with Homebrew first if needed; the package manager on Linux), Copy and open Terminal on this Mac, and Check again. Nothing is created until it's there.
v0.3.4
October 5, 2026.
Two worktrees side by side, nineteen themes, a chat that keeps every turn, and browsers that show you the page.
- Fix a blocked agent's browser from Berth. On Ubuntu 24.04, whose
sandbox setting stops Chromium starting, the browser pane marks the
agent's browser blocked and Settings → Boxes shows why, with two ways
out: Fix it opens a terminal on the box with the
sudo sysctlcommand typed for you to run (Berth never sees your password) and clears once the box allows it, or Run without Chromium's sandbox, a per-box setting that needs no restart. An agent that hits it is told to ask you, andberthd doctorsays the same. Boxes need an upgrade. Agent browser. - The browser tab is your browser. A browser pane always shows the
worktree's page in a real browser on your laptop, through its proxy; the
agent's browser on the box shows only while an agent drives it, behind a
clearly marked Agent's view · live, and now shows the page at once
rather than after its next repaint (most still pages never drew). URLs
the box writes with its own name (
$BERTH_URL,devboxwhere you paired it asdevl) open too, a redirect to another port on the box'slocalhostgoes to that port, and a server macOS can't name counts as the worktree's dev server again. Boxes and the laptop agent need an upgrade. Agent browser. - The chat keeps every turn. A prompt with something pasted in it
(Claude Code 2.1 writes those with a new closing tag) was left out of the
chat entirely; it now shows as typed. A chat also lost or jumbled turns
when the box read the conversation afresh (after a box upgrade, or once
no one had looked at it for ten minutes) or a long chat let its oldest
live turns go while older ones were loaded: items are now named by where
they are in the agent's record, so the chat keeps what it showed, reads
what it missed in between, and turns read back after being hidden come
back on their own.
/compactreads as one divider. Boxes need an upgrade. - Worktrees side by side (Labs). Two or more worktrees share one tab strip as tab groups: ⌥-click one in the sidebar, drag it onto the strip, or choose "Add to tabs beside…". Each group has its name and colour; click it to fold it, right-click to colour or close it (its agents keep running). ⌃⇧⇥ and ⌃⌥⇥ step between groups, ⌘1–9 count across the strip and ⌘⇧W closes the group in front (with one, the window). A tab can show panes of two worktrees too: ⌘⌥D picks a worktree, then one of its agents, its page, its diff or a new terminal, and puts it beside the focused pane; dragging a worktree onto a pane's edge splits its agent in. Run, ⌘T, ⌘⇧B, the + menu, the palette and plugins act in the focused pane's worktree. With one worktree open nothing looks different. More.
- Scrolling keeps up. The wheel over a terminal now reaches the agent in batches, one per redraw, rather than one report per wheel event. A trackpad flick through a long Claude Code thread queued dozens of full-screen redraws behind each other, and over a slow connection to the box the view kept scrolling for seconds after you stopped; now it stops within a redraw, the same distance down. Drawing a whole new screen in the terminal also takes half the time it did.
- Nineteen more themes. Ports of popular VS Code themes: Dracula, One Dark Pro and One Light, Catppuccin Mocha and Latte, Tokyo Night, Nord, GitHub Dark, Dark Dimmed and Light, Gruvbox, Solarized, Rosé Pine and Dawn, Night Owl and Night Owl Light, and Kanagawa Wave, each from its own palette for the app, the terminal and code, nudged only where text would be too faint to read. Settings groups them Berth, Dark and Light, with a search, and Match system can use any light and dark pair. Themes lists them and how to make your own.
- Code in replies is coloured. A code block in an agent's reply is now syntax-highlighted like the app's diffs, in the theme's own colours, by its fence's language (or a guess). It colours in after it draws, off the main thread, without moving; Copy still takes the plain code.
- Smaller fixes. A long reply scrolls inside the reply box instead of pushing the chat away, and text size (⌘+ / ⌘−) now reaches the chat itself, its code and its edits, not just the UI around it.
v0.3.3
October 5, 2026.
The chat takes on the rest of what a terminal does: Claude's questions, its artifacts, its edits as diffs, and dev servers in tabs of their own.
- Archiving follows through. Archive or remove a worktree from the sidebar, zen, the launcher or the Worktrees table and it shows as "Archiving…" (dimmed, nothing to open) while the repo's archive script runs, then leaves the sidebar, the table and its open tabs as soon as the box reports it gone; its sessions stop with it. If the script fails, the worktree comes back with the script's output under Details. The app used to say "Archived" at once and keep showing it. Boxes need an upgrade for the session stop and the script's output.
- A worktree left locked archives. A
git worktree addthat was cut off leaves git's own "initializing" lock, and git then refuses to remove the worktree. The box now unlocks and removes those, cleans up after an add that fails, and still leaves alone a worktree a person locked, saying how to unlock it. Boxes need an upgrade for it. - Details opens. "Details" on an error toast did nothing when clicked. It now opens the full output (git's, a script's) in a dialog with Copy, everywhere it appears.
- Artifacts in the chat. A page Claude publishes on claude.ai with its Artifact tool shows as a card where it happened, with Open, and a "3 artifacts" chip at the right of the reply box's controls lists them all, newest first: open one, copy its link, or show it in the chat. A page published again is marked updated. Boxes need an upgrade for it.
- Claude's questions, answered in the chat. When Claude asks several questions at once (AskUserQuestion), the chat used to answer only the first and then hang. Now it shows the whole form: each question with its options, one pick or several, room for your own words, back and next, and a review with Submit. The box fills in Claude's own form key by key, checking its screen as it goes; if anything doesn't match, it stops and Claude's screen opens under the chat to finish there. Answers stay in the conversation as what you picked. Boxes need an upgrade for it.
- Services in a terminal tab. A service with
"terminal": true(in a kit,.berth/config.jsonor a box's own config, with an optional"title") runs in a terminal of its own, which opens as a tab named after it when its worktree does, so a dev server's output is right there. Ctrl-C in the tab stops it, and Start, in the tab or anywhere else, runs it again in the same tab. Closing the tab only hides it; "Running in" and the Run menu have Show terminal. Boxes need an upgrade for it. - Drag tabs into splits. Drag a tab from the top bar over a pane and drop it on an edge to split that pane right, left, up or down, with an outline showing where it lands; drop it between tabs to reorder. A pane drags by its header beside another pane, onto another's middle to swap them, or onto the strip to become a tab of its own. Nothing restarts: terminals and agents keep their sessions. A tab's menu has Split right and Split down, and a pane's ⋯ menu has Move to a new tab.
- Edits in the chat look like the Diff panel. An agent's edit opens highlighted, in the same renderer as the Diff panel, numbered by the file's own lines; Write and MultiEdit too, and the file's uncommitted diff beside it still takes comments. Unified in the chat's column, split where the pane has room (it remembers), and a long edit folds to Show all. The renderer loads the first time you open an edit, never before. Boxes need an upgrade for the line numbers; older ones count from 1.
- Chat backgrounds and width. Settings → Appearance → Chat background puts a quiet background behind conversations, in zen too, drawn in your theme's colours: patterns (chart contours, a dot grid, a nautical grid, soft waves), dithered gradients (Dawn, Dusk, Deep sea) and harbour scenes. A picture of your own, picked, dropped or generated with Codex on your Mac (it shows the exact command and runs only when you press Generate), lands dithered in the theme's ink. Strength, Dither and Tone shape it; Advanced can show a picture as it is. Chat → Width sets how wide conversations run, from Narrow to Full. A pane's ⋯ menu has Chat background….
- One corner radius. Cards in the chat, the task list, notices and the agent's own screen take the reply box's radius; bubbles keep their own.
- Zoom with ⌘+ and ⌘−. View → Zoom in (⌘+ or ⌘=), Zoom out (⌘−) and Actual size (⌘0) make the app's text bigger or smaller, from 11 to 20px, with a moment's readout of the size; it's kept, and Settings → Appearance → Text size shows it. With a terminal focused they size the terminal's font instead, and ⌘0 puts it back to 13px.
v0.3.2
October 5, 2026.
- Built-in plugins on a fresh install. The release's content policy kept the app from reading its own list of built-in plugins, so a new install had none: Diff, Changes, Pull request, Notes and the rest. Dev builds skip that policy; CI now checks it.
- New task with a line break (⌘N) starts its agent. The box couldn't read back a session whose prompt had a line break, so it removed the worktree it had just made and said "It's gone".
- Pasted screenshots upload fast. A pasted image over 1 MB or 2000 px is shrunk first (a 4 MB screenshot goes up as about 30 KB), files go up as raw bytes rather than base64, and the chip shows progress, its size, and Retry if it fails. Chips sit inside the composer.
- Replies with a paste show where they were sent. Claude Code records a
paste wrapped in
<pasted_content>tags; the chat dropped such a reply, and it showed below the answer instead. Its tags no longer show either. - A reply sent while the agent works says it hasn't read it yet: "Sent · Claude reads it at its next step".
- Diff and Pull request follow the branch when the agent switches, commits or rebases mid-turn.
- Live text read off the agent's screen looks like the reply, not a quote.
v0.3.1
October 5, 2026.
- Labs is on by default: the harbour home, agents as conversations, and zen. Settings → Labs turns it off, and that choice is kept.
- Claude Code's idle prompt no longer reads as one of its own screens,
which opened the live terminal over an empty chat (Claude Code 2.1 draws
a no-break space after
❯). - A reply in a long chat no longer says the agent "hasn't taken this yet" or shows twice.
- Edits that only add lines don't show a stray "−".
v0.3.0
October 4, 2026.
The chat grows up: it stays in step with the agent's terminal, takes everything the terminal takes, and stays fast on long conversations.
-
One chat per agent, in step with its terminal. Several agents in one folder keep separate chats; a new agent never shows another's. A message sent mid-turn shows before its reply, the running step and the agent's own status ("Running pnpm test · 18s", "Seasoning… · 9m") show live, and its words show from the screen before the transcript has them.
-
Replies as Markdown. Headings, tables and code render as the terminal shows them, with less noise: the steps between replies fold. Text selects and copies; Copy takes the Markdown.
-
Open a tool call to see what the terminal shows: the full command and its output, an edit's exact change (
GET /v1/sessions/{name}/transcript/tool/{id}). -
Everything the terminal takes.
/lists the agent's built-in and custom commands, skills and plugins;@files,!shell. Screens of the agent's own, such as/model's picker, open inline and fold away. -
Chat controls. Stop, the permission mode and model chips, the agent's task list, context meter, background shells, and cards for API errors, limits and an agent that exited mid-work.
-
History and scale. Older turns load as you scroll up, helpers' (subagents') own chats open from Crew, ↑/↓ recalls prompts, ⌘F searches the chat, and Fork and Rewind start over from any prompt. 5,000 items paint in about 0.3 s and scroll at 60 fps.
-
Paste or drop images and files for an agent on a remote box, in the chat or its terminal; they upload to the worktree's
.berth/attachments. -
What's running in a worktree sits under the composer, with names and Open in tab, and plugins can add a section with
addWorktreeSection. -
Diff panel. A worktree's whole branch diff, its uncommitted changes, or both, drawn with
@pierre/diffs. -
Pull request picks. Pick failing checks, reviews and comments on a PR and send them to the worktree's agent, or copy them.
-
Closing a tab stops its agent, with a moment to undo, and a new tab is always a new session.
-
Zen on ⌘. from the menu bar, rename and archive from zen, and the wheel scrolls a full-screen agent's terminal instead of cycling its prompts.
-
What happens to your message. A reply sent while the agent works shows as queued in the conversation, with Send now and Cancel, and the reply box says whether Enter sends or queues. The box lists, cancels and sends held prompts (
GET/DELETE /v1/sessions/{name}/queue). -
Approvals from Claude Code's hooks. A permission reads "Claude wants to run
pnpm test" with Claude's reason, from itsPermissionRequesthook rather than the screen, with Allow, Always allow and Deny matched to its menu. The request stays in the box's private ledger, never in events. -
See the work, comment on it. An edit in the conversation opens to its diff (
GET /v1/sessions/{name}/diff), and lines take comments there and in Review, sent to the agent as one short prompt.
v0.2.0
October 4, 2026.
Orchestration rebuilt on the box: it keeps going while your laptop sleeps, survives berthd restarts and upgrades, and costs your agents as few tokens as it can.
- Turns, not guesses. Every prompt Berth sends gets a turn ID, and
waits, loops and flows wait on that turn, not on the clock or the folder.
Each session is tracked on its own, so a reviewer beside an agent no
longer ends the agent's wait. Hook events go through a numbered journal
(
GET /v1/events?since=), are kept while berthd is down, and are never dropped in a burst. Codex, Cursor, Gemini and OpenCode report their state through their own hooks. See sessions and agents. - Quieter "needs you". Claude's idle reminder no longer counts as a question, an approved permission clears "waiting", and nothing Berth sends (broadcast, the queue, loops) types into an agent that is waiting on you.
- Durable runs. Loops, reviews, hand-offs, broadcasts and flows run on
the box from a journal: they keep going while the laptop sleeps and
resume after berthd restarts, never sending an agent a prompt twice.
berth run …,berth runs, andberth loopas a run; gates you decide from the app, the CLI or the phone; box-wide caps and per-run budgets. See runs. - Try N ways. Several agents, or models, try one task in their own worktrees, on one box or several; a check verifies them, a headless judge ranks them, and Review's Compare lets you pick one for a draft PR.
- Triggers v2. Issue labels and assignees, signed webhooks, and an Auto-fix this PR toggle; a GitHub item arriving while a flow runs waits for it instead of being dropped, and comment text reaches an agent labelled as untrusted, from the authors you allow. See automations.
- Handoff packets and native fork or resume; headless agents in
tmux;
berthd mcpfor agents. The app's flows for every box are gone: commit such a flow to the repository's.berth/config.jsonor its kit. - A browser for your agents. Each worktree's agent can open its own dev
server at the URL you see (
BERTH_URL) in a headless Chromium on the box, confined to that worktree's ports, and act on it in a few hundred tokens a step. Watch it live, pick an element to send it, and see its screenshots in Review. On Ubuntu 24.04, allow Chromium's sandbox or opt out withBERTH_BROWSER_NO_SANDBOX=1. See agent browser. - The agent's conversation, read from Claude Code's or Codex's own
transcript (
GET /v1/sessions/{name}/transcript), for a calmer view of what an agent is doing. Tool output and thinking are never sent. On a box with an older berthd, the view offers to update it in place. - Pick a model and effort for a new agent (
modelandefforton tasks and attempts, passed to each agent's own flags), so attempts can pit models against each other. - Labs: a calmer Berth. Behind a Labs switch in Settings: a harbour home with one composer and your agents, a Conversation view beside each agent's terminal, and zen (⌘.), which puts the sidebar away. See Labs.
v0.1.2
October 3, 2026.
- Set up a box from your tailnet. Add a box lists the machines on
this Mac's tailnet, with their OS, whether they are online and which are
boxes already, and Set up installs and pairs one over SSH. Tailscale
SSH needs no keys, a host key the tailnet vouches for is trusted, and
check-mode approval opens in your browser. Sign in to another tailnet
lists that tailnet's machines too.
berth discover --jsonreports Tailscale's state and each machine's Tailscale SSH and host keys. See add a box. - A kit can live in a repository of its own.
Kit links to GitHub, GitLab, Bitbucket and Codeberg work without
https://(berth kit add github.com/me/kit). See kits. - Add another computer with a join link, and use this Mac as a box without the CLI.
v0.1.1
October 3, 2026.
- A signed Mac app.
Berth-macos-universal.dmg, one app for Apple silicon and Intel, signed with a Developer ID and notarized. It carries theberthCLI and the Linux daemons. See install. - Automatic updates. The app checks for a new release in the background and shows Restart to update when one is ready; it never restarts by itself.
- The
berthcommand from the app. Settings → General → Command line links~/.local/bin/berthto the app's copy. - Security fixes from an audit; the security model
has the whole picture:
- A repository's committed
.berth/config.jsonruns on a box only once trusted there, for its exact content. - Plugins you install never turn on by themselves, and ask for consent pinned to a hash of their code.
- Flow values reach
shell steps as
$BERTH_FLOW_*environment variables, never as script text; webhook values are encoded; secrets show as[secret]. - Webhooks and phone notifications can't reach loopback, private or
metadata addresses unless
~/.berth/network.jsonon the box allows them. - Gates cover every route that changes something and see an origin the box can vouch for; revoking a laptop closes its open streams; pairing is limited per source address; clones and kit fetches refuse local and command transports; kits apply only what you reviewed; project skill installs refuse symbolic links.
- A repository's committed
v0.1.0
October 3, 2026. The first release.
berthdfor Linux and macOS boxes, and theberthCLI, for amd64 and arm64, withchecksums.txt.- One line installs berthd on a box and prints a pairing link:
curl -fsSL https://berthd.app/install | sh. It installs the hooks that let Claude Code, Codex and Cursor report their state. berth add sshinstalls and pairs a box over SSH in one step.- The desktop app, built from source, with the laptop agent startable from the app.