What changed, and why.

New

• The App librarian. Every App now has a librarian, switched on by default: an agent that keeps the App's knowledge current. On each run it:

  • shares the App's code, deploy and decision docs with the App;
  • marks older and duplicate docs as superseded;
  • keeps one " docs index" doc;
  • writes the AgentBook pages Deploy and roll back, Environments and access (where things live, never secrets) and Local setup and testing;
  • raises one task for anything nobody has documented.

  It runs weekly, after 5 new docs, and when a workspace on the App is deployed, changes stage, is completed or stopped, with at most one automatic run every 6 hours. It runs in a read-only box on the App, so it never changes code. Run now, the on/off switch and its recent runs are on the App's Settings tab.

Changed

• Docs are shared with the App by default. A doc written in a workspace that belongs to an App is listed in every workspace on it and in their APP.md. devbench doc create … --workspace-only keeps a working note to its workspace. devbench doc share <id> shares an existing doc, and devbench app docs lists every doc on the App.
• Docs outlive their workspace. Archiving or deleting a workspace leaves its docs with its App.
• One current version. A doc shared with an App under the same title as an older one there supersedes it, so the App shows one runbook, not two.

New

• The chat answers when the agent is quiet. If the agent has said nothing a minute after your message, DevBench posts one line for it, quoting the agent's own latest progress: "Claude has been working on your message for 3 minutes and has not replied yet. Latest: "Server sign-in tests pass (14). Next: the push routes.""

  • The line updates in place while the agent works.
  • It disappears when the agent replies.

• If the agent stops without replying, DevBench asks it once to answer. If it still doesn't, or its session is down, signed out or out of usage, the chat says so and the team is alerted.

• Each message says what happened to it, under the time:

  • Sending to Claude…
  • Delivered to Claude
  • Held until the question is answered
  • Not sent to Claude (read as talk between people: mention @agent to send it)
  • Didn't reach Claude

  This shows on the web, the desktop app and the phone.

Fixed

• A message is no longer lost when DevBench restarts while it is on its way to the agent, for example while a stopped project wakes up. It is sent again. If that fails too, the chat says to send it again and the team is alerted.
• Agents are reminded mid-turn. In projects with the slim agent instructions, an agent that has worked 90 seconds on a chat message without posting is told to post a one-line update. Before, the reminder came only when its turn ended.

New

• Realistic typing. Each field now gets a few realistic values to choose from instead of one generic test value, and Jev predicts which one a real user would type there. The values come from three places:

  • the App's sample values;
  • the screen itself: the field's own example, and items already listed, so a new project is called "Website relaunch 2" rather than "Explorer Test";
  • what the field is: names, companies, cities, postcodes, prices, titles, messages and more.

  Mistakes (empty, the wrong format, too long, odd characters) are still tried on every form.

• Letter by letter, with the app's suggestions. The walk types the way a person does. When an autocomplete or search list appears partway through, Jev picks the suggestion the user would choose, or keeps typing. If picking a suggestion leaves the field empty, that is reported as Choosing a suggestion did nothing.

• What the explorer types, in an App's deployment defaults: sample values as Field = value lines, such as Project name = Website relaunch. These are offered first.

Changed

• When the new links run out, a walk now goes back to screens that still have controls it hasn't tried, such as a form it left for a navigation link, instead of stopping with steps left.

New

• The explorer signs in by itself. Save EXPLORER_LOGIN_URL (a link that signs in, such as a demo user's login-token link), or EXPLORER_EMAIL and EXPLORER_PASSWORD for a test account, as workspace secrets. Paste NAME=value in the workspace chat, or use Settings › Secrets. Every walk then signs in first, whether started from Explore now, after a deploy or with devbench explore. The values are read by name in the box and hidden in its output. The Previews tab and the phone's Deploy screen say how walks sign in, and what to set when nothing is.

Changed

• Setup problems come first, apart from the app's bugs.
  • A preview part that isn't running is the walk's first finding, with its last log line. For example, a backend that exits on start-up because its start command runs a command that doesn't exist.
  • A walk that never gets past the sign-in page reports Stuck at sign-in and names the secrets that would let it in. If it was given a way in, it says the link or account may have expired.
  • The estimate sent to the agent lists these first as setup problems with the preview, not bugs in its change.

New

• Explore it, from the chat. When a preview comes up and isn't explored automatically, or a version is deployed to an environment, the workspace chat asks Explore it / Not now. For staging and production it warns that forms the walk fills in are submitted for real. A person answers. Explore it starts the walk, and its estimate follows in the chat.
• Agents can ask you to run it. devbench explore ask "<why>" posts the same card, with the agent's reason. An agent cannot answer it itself.
• The explorer in the phone app. The Deploy screen has an AgentReview Explorer section: Explore now, the after-every-preview choice (as the App says, always, or only when asked), and recent walks. Tap a walk to see its findings with the steps to reproduce them, open its report, or stop it. Settings has the same switch.
• Android builds on Lathe phones. devbench explore --device <phone> --package <app id> walks an installed Android build the way the web walk does:
  • each screen is read from the phone's element tree;
  • taps, typing and Back go through Lathe, so the phone holder's switches decide what it may do;
  • a crash is the app's fatal error in the logs, or the app leaving the screen. The walk then reopens the app and carries on.

Changed

• A walk's report link is included in the API (report_link), so the phone can open it.

Added

• Invite codes are single use by default. A new code lets one person in. Whoever makes it can choose more people, or no limit, in the Invite dialog, and change it later from the code's row on the web or its menu on the phone. Every code that had no limit is now single use; one that had already been used is now used up.
• The message box shows formatting, not Markdown (web and desktop). Bold is bold, a list is a list, code looks like code, as the message will read once sent. It is still sent as Markdown, as the agent and everyone else expect.
  • Markdown still works as you type: **bold**, *italic*, `code`, ~~strike~~, - or 1. for a list, > for a quote and three backticks for a code block turn into their formatting. Markdown pasted as plain text becomes its formatting too.
  • Shift+Enter is a new line (a new item in a list; on an empty item it leaves the list). Enter sends, or Ctrl/Cmd+Enter if that is your setting. @ mentions work as before.
  • What is sent is what you typed: no backslashes are added before a * or [ that would not have formatted anything (a key, a path, snake_case).
• A formatting bar on top of the message box: bold, italic, strikethrough, inline code, link, bulleted, numbered and check lists, quote and code block, with their shortcuts (Ctrl/Cmd+B, I, E, Shift+S; Shift+7, 8, 9 for lists; Shift+B for a quote; Alt+C for a code block). At its right, Show Markdown switches the box to the Markdown underneath and back; the choice is remembered in this browser.
• On the phone, long-press text in the message field to format it, as in WhatsApp: a menu comes up over the field with Bold, Italic, Strike and Mono, and More for a link, lists, a quote and a code block. It puts the Markdown in for you (the phone shows the Markdown as typed for now). Nothing is selected, nothing shows.
• Pop out videos and docs from chat, as photos open. A video has a pop-out button beside Download (on its poster and while it plays), and a DevBench doc attached to a message has one beside its link: both open in the viewer, where the arrows walk a message's photos, videos and docs together. The doc reads right there.
• The viewer covers the workspace only. The top bar and the rows of project tabs and groups stay usable while a photo, video or doc is open, so you can switch to another project or group without closing it first. It has Open in a new tab beside Download.
• Short, tidy addresses. The address bar no longer shows the database's numbers: a project is /workspaces/0mVgX2, a doc /docs/3JTTPW, and a group is one short link, /g/1fJ2h, in place of /workspaces/289/?tab=chat&beside=288,414. A group's link carries its projects, so a teammate who hasn't saved that group still opens the same projects side by side. Copied links (a task's, a message's) use the short form too, and old links with numbers still open.
• Merge requests: Go to chat and Ask agent to review. Every waiting merge request, on the phone and the web, goes to its project's chat in one tap, and Agent review asks the project's agent there (@agent, in the box that made the change) to review the whole change, run the tests, fix what it finds and merge it when it's clean. It asks you first, then opens the chat so you can watch.
• Working now on Home shows the agent's logo, not its name. The row is the project's name alone, with Claude's (or Codex's, OpenCode's) mark beside it, so long project names fit.

Fixed

• Joining with an invite code no longer loops back to the code page. Continue opened WorkOS's sign-up screen, which in our WorkOS setup sends people straight back to /signup, so it looked like a page refresh with nothing happening. Joining now opens the sign-in screen, where Google, GitHub or a code by email creates the account. If someone does come back without signing in, the page keeps their code, says what to do, and the team gets an alert when they try again.
• Copying a chat message's numbered list and pasting it nested the list one level deeper each time ("1. 1. 1."). The list's numbers are drawn beside each item, and a copy took them as text; they are no longer copied, and the paste ignores them. Copying a message also keeps its bold, italics and strikethrough now.
• A public doc link stays public when you're signed in. Opening a shared doc link (/d/…) no longer moves you into DevBench; you see the page anyone else sees, with Open in DevBench at the top right when the doc is your organization's.
• Links, everywhere (from the link review):
  • A link written without https:// ([site](example.com)) works; docs/12 or README.md stays text, and //other-site links are refused.
  • A link around code ([`README.md`](…)) stays a link, and an address right before **bold** or a backtick stops at the address.
  • A number in rendered Markdown (task notes, Inbox, docs, MR notes) is no longer drawn as one of the doc's links, and a link with mixed bold and plain text is one link, not several.
  • The desktop app hands only web and mail links to the computer; anything else a page asks it to open is refused.
  • The phone opens DevBench doc links in the app, keeps the task or doc a link points at, opens web links in its own browser, says so when a link can't be opened, asks before following a link whose text shows a different site, and offers Open / Copy link / Share on a long press.
  • A message's Copy link on the phone gives an https:// link; old devbench:// links open the message on the web and the phone too.
  • The link dialog makes an email address a mailto: link and accepts localhost:3000; clicking a desktop notification opens the message itself; long links wrap with the sentence.
• Ctrl/Cmd-click on a group opens it in a new browser tab and leaves this page alone. It used to move the page you were on as well. The same goes for every control in the top bar that moves the page through the app (the project switcher, All projects, a project's Settings): a Ctrl/Cmd-click opens a new tab and nothing else.
• Phone: videos play full screen, and are drawn in their own shape. A video in a chat message, a task note or a doc has a full-screen button beside its play button, which opens the phone's own full-screen player (playing, with its controls, turning with the phone). A portrait recording is drawn tall and whole instead of cropped into a 16:9 strip. The merge request and Files players allow full screen too.

Changed

• Organization is a tab in the top bar, last, after Insights (Home, from 0.4.81, is first). The menu under the organization's name still works, and Alt+Shift+9 opens it, as Alt+Shift+1…8 open the other tabs.
• The Organization area, redesigned: seven sections in one layout. Each section has a heading with a mark, its settings in grouped cards, and fields that save as you change them (no Save bar; the OpenAI key has its own Save so a half-typed key is never stored).
  • General: rename the organization (new), the plan and the compute used this month, export all of its data (new) and delete the organization (new: type its name to confirm; it is deleted after a grace period and can be cancelled until then).
  • Members (was Team): people, roles, suspending, invites.
  • Agents: agent accounts and failover, the agents' versions, the agent and model new projects start with.
  • Agent behaviour: team instructions, which chat messages reach the agent, the conversation history, the task checker, the OpenAI key for drafting. (Agents and Agent behaviour were one 2,700 px page, AI.)
  • Projects: the setup script, board stages, git flow and protected branches, and the boxes' CLI.
  • Integrations: GitHub, GitLab, Bitbucket, the inbound webhook, and API keys (were a page of their own).
  • System: server health (was a page of its own), backup and restore, and a way to Super admin for super admins.
  • Old addresses still work: /team, /api-keys, /monitor and every old ?tab= open the section their settings moved to.
• Pages use the width of the screen. List pages (Inbox, Tasks for you, Merge requests, Docs, Agents, Schedule, Watchers, Tools, Apps, Cloud, Releases, Previews, every Lathe page including its setup, a project's Overview, Docs and Activity) were capped at 1,024 px in the middle of the screen, and pages of prose at 768 px; on a 1,680 px screen a third of it on each side was empty. Lists and settings now use up to 1,440 px, prose up to 1,024 px, and Install DevBench and Preferences the list width. Project settings' column beside its rail is wider to match.

New

• See what a project box on a Lathe machine is doing. It used to read only "provisioning" or "starting" until it was ready. Now the project's panel, Settings → General and the machine's tile in Lathe → Machines show:
  • which machine it is on and what that machine is doing: downloading the image recipe, building the box image (with the step, e.g. "step 12 of 30", and the instruction it is on), or creating the box;
  • how long that has taken, and when the build last printed anything, with a warning when it has been quiet for over five minutes;
  • once the box runs, DevBench's own setup step, e.g. "Cloning devmustafa4/vt-linux". This needs Lathe for PC 0.3.3 on the machine; an older one still reads "Starting on its machine".
• Lathe → Machines lists the project boxes each machine runs. A machine setting one up reads "Online · setting up a project box", not "ready".

Fixed

• A machine's first box image build no longer fails the project at 45 minutes. The 45 minutes now count from when the machine has the image. The build itself has up to 3 hours, and one that prints nothing for 35 minutes is stopped with the step it was stuck on.

Changed

• Five sections instead of nine. Project settings had nine sections, three of them holding a single setting, with the agent's model under Automation, the box's switches split across General, Automation and Security, and Archive and Delete in different places. Now:
  • Box: start or stop it, its memory, where it runs, automatic shutdown, AWS access.
  • Agent: its model, its Claude account, the config template, and the rules the box holds it to.
  • Code & merging: self-merge, Second Look, push guardrails, the release flow, reviewer comments.
  • Secrets & access: secrets, SSH servers, the production alert token, the inbound webhook.
  • Project: name, stage and deadline (on Overview until now), Archive, and Delete last, in red.
• One row each, the detail in a modal. Each setting is a title, one line and a switch or a button; what takes room to change (the model picker, the agent rules, guardrails, the release flow, secrets, servers, the alert token, the webhook, renaming) opens in a modal.
• Everything saves as you change it. Agent rules used to wait for a "Save agent rules" button at the bottom; a switch now saves at once, and a list or a command when you leave the field.
• Owners can show a secret's value from Settings too, as from its 🔒 name in the chat (audited).
• The phone has the same five groups, and gains what it lacked: Second Look, the agent rules (switched on and off there), and the project's secrets (owners can show a value).
• Old links such as ?section=merging or ?section=danger still open the section those settings moved to.

Added

• Box memory in project Settings. Every box used to get the server's one size, 6 GB. In Settings → General (and Settings → Box on the phone), an owner can now choose 2, 4, 6, 8, 12, 16, 24 or 32 GB for the project, or the server's default. The row shows what the running box uses of what it has.
  • Raising it reaches a running box at once, with no restart.
  • Lowering it does too when the box uses well under the new size. Otherwise it is saved and applies at the box's next start, so the agent is not squeezed out mid-run, and you are told which.
  • Owners only, signed in as themselves: the box's own key cannot raise its own ceiling. Each change is in the project's audit log, and a change that fails on the server alerts the team.
  • Linux boxes. A Windows box's memory is its VM's and is not changed here yet.
  • The most a project may ask for is DEVBENCH_WORKSPACE_MEM_MAX_GB (default 32).

New

• Files over 512 MB, with a person's yes.
  • An agent runs devbench media request <file> "<why>". That posts an Approve / Deny card in the project chat with the file's name, its size and the reason.
  • An organization owner, or anyone who manages deployments, can approve. An approval lasts 24 hours, for that one file at that exact size.
  • devbench media put then sends the file straight to S3 in 64 MB parts. It never passes through DevBench's server and is never held in memory. A failed part is retried, and running put again picks up where it stopped. The result is one file with the usual share link.
  • The parts are checked before the file is finished. Half-finished uploads are cleaned up after a day. Up to 50 GB.
  • devbench media requests lists an agent's requests, and devbench media abort <id> gives one up.

Fixed

• An agent can no longer approve its own request. An agent's key acts as the person who made it, often someone who may approve. Its reply in the chat ("1"), or its answer through the API, used to count as that person's yes on its own access request. Approve/Deny cards for server access, cloud access, destructive calls and large uploads are now answered only by a person.

Added

• Press a secret in the chat. A key pasted into the chat is saved as a project secret and the message keeps only [🔒 NAME]. That name is now a chip, on the web, the desktop app and the phone. Pressing it opens:
  • Show value: the value, shown to you in that dialog only. It is for the organization's owners, signed in as themselves: no API key gets it, the agent's included, and the agent never receives it from here. Each view is in the project's audit log, and the answer is never cached.
  • Rename: the new name, with every marker in the chat following it.
  • Not a secret: for a value hidden by mistake. After asking, it goes back into the chat as it was written, everyone (the agent included) can read it, and it is removed from the project's secrets. The chat says who did it, by name only.
• A marker inside inline code is a chip too, so DevBench's own "In code it is NAME=[🔒 NAME]" line can be pressed.

Changed

• No more polling in the desktop app. The desktop window talks to a small server inside the app, which passed requests on but not the live connection, so the window never went live: it kept every timer the web app switched off in 0.4.66, and its chats polled too. The live connection now goes through to the server, so the desktop window behaves like the web app.
• No more polling in the phone app. The phone now keeps one live connection while it is open and someone is signed in, and reads only what the server says changed. These stop asking while it is up:
  • the tab bar's counts (every 45 seconds);
  • a project's page (every 6 seconds while its agent works);
  • the agent screen's state and questions (every 10 seconds);
  • the agent transcript (every 4 seconds; it now follows the box's own socket, as the web app does);
  • a chat's questions and @agent relays (every 10 and 8 seconds).
• The phone follows changes it used to show only on the next visit: Home, Agents, the board, Activity and a project's tasks update as the server announces them.
• The server also announces agents' questions (asked, answered or withdrawn) and @agent relays.

Fixed

• "The finish line could not be handed to the agent" no longer fires for a finish line that landed. A finish line is sent to Claude as one long line, and Claude Code shows a long single-line paste as [Pasted text #N] with no line count. DevBench only recognised the counted form, so it judged every long finish line undelivered, cleared it and said so in the chat. It now recognises the uncounted form on the input line. The same applied to any long single-line message to the agent.
• In the background the phone closes its live connection, so it costs no battery; back in front it reconnects and every open screen reads once to catch up. If the connection drops, each screen polls as before until it is back.
• Pages that follow a running job (a deploy, a config scan, a backup, logs, Lathe) still poll while it runs, as on the web.

Changed

• Groups you can see at a glance. Projects open side by side are a group, and each group is now drawn as one box, a wash of its colour round its head and its tabs (no coloured outline, which 0.4.61 took off the tabs; the group on the screen gets a hairline, as a selected tab does). A project outside any group sits on its own, with no box.
• A group's head has three parts:
  • The icon on the left opens the group's settings: open side by side, name, colour, collapse and ungroup. Right-click still opens them too.
  • The name and how many projects the group holds. Click the name to rename the group right there: Enter saves, Escape keeps the old name, and an empty name removes it. A group with no name says "Group".
  • The arrow points left to collapse the group to its head, and right to expand it again. Alt-click on the name still does the same.
• Opening a group side by side is now the first item of its settings, or a click on any of its tabs. A click on the name no longer opens it, because it renames it.

Changed

• One-handed phone app. The actions people use most moved to the bottom third of the screen, where the thumb already is:
  • A project shows Chat · Agent · Tasks in a bar pinned at the bottom, with the unread count on Chat.
  • Chat and the agent screen have a Project · Chat · Agent switch just above the message box, so you can go between them in one tap. On the agent screen the Conversation / Terminal switch, the conversation picker and the terminal's keys are down there too.
  • On the Projects board, Groups, Filter (with the stage), Search and View sit in a dock above the tab bar. Groups open as a sheet with large rows and a chat button on each project. Every card has Chat and Agent buttons along its bottom edge, and a long press opens the project's menu.
  • Chats, Inbox and More have their search and switches at the bottom as well. A Tasks list has its Open / Done / All switch and an Auto-run button next to the box where you add a task.
  • Every "+" (New project, New request, New doc, Upload, Invite and the rest) is a round button at the bottom right instead of a header corner.
  • Pop-up sheets put Close, Back and Reset at the bottom, and tall ones can be swiped down. Search fields in sheets and on the Search screen sit right above the keyboard.

Fixed

• The phone app is faster.
  • It opens straight onto the board and chat list it showed last time and updates them in place, instead of waiting for the server before drawing anything.
  • Screens you return to show what they had at once.
  • The same request is no longer sent twice at the same time, and the screen's own requests go out before the badge counts and other background reads.
  • The tab badges read a count instead of downloading whole lists, and stop while the app is in the background.
  • The Inbox loads its tasks in one request instead of a chain of them.
  • Typing in a chat no longer redraws every message, and background checks that find nothing new no longer redraw the thread.
  • Reconnecting no longer re-formats every message, and the agent's terminal no longer redraws the whole screen many times a second.
  • The Tasks and Logs lists only redraw the rows that changed.
• Slow launches and slow requests on the phone are now reported to the team (mobile.perf.slow-launch, mobile.perf.slow-request).

New

• Live updates. When a project, one of its tasks or its chat changes, every DevBench window open in the organization hears about it at once, over the connection it already keeps for notifications, and shows the change straight away. The project board used to ask for the whole project list every 4 seconds, and the project tabs every 20; while the connection is up they now ask only when told something changed, plus once a minute as a safety net. When the connection drops they go back to asking as often as before, so a lost connection costs freshness, never correctness.

Changed

• A much smaller first download. The code the browser downloads before it can draw anything went from 1,136 KB to 556 KB, and opening the project board fetches 36 files (about 870 KB) instead of 78 (about 2 MB): the project page, the code highlighter, the release notes, the brand icons and the sign-in pages now load when they are needed, and the project page loads quietly in the background once the board is up. The app's files are kept by the browser for good, so a second visit downloads only what changed in a release.
• Data from the first moment. The page asks who is signed in, for the projects, and for the project in the address as soon as it opens, while the app's code is still downloading, and starts fetching the code of the page in the address at the same time. The first data used to be asked for 1.4 to 2.2 seconds into a load; now it is asked for at 0.1 s. Requests the app made several times at once (the project list four times while the board opened) are made once.
• Faster answers from the server. The project list does far less work per project and per chat message (about 0.25 s to 0.1 s on a board of 118 projects). A project's git status, the agents' working state and the merge-request queue answer at once from a short-lived shared copy and refresh it in the background; Refresh still asks the box there and then, and a push or merge from DevBench is shown straight away.
• Calm loading. A page whose code or data is still on its way shows a placeholder shaped like the page, and only if it takes longer than a fifth of a second, so a quick load shows nothing at all; the "Loading…" lines are gone. Hover and press feedback is quicker and the same everywhere, and nothing moves under the pointer. Empty board columns say "No projects".
• Monitoring: the early requests of a page load are logged, and the team is told if they all fail; the server logs its shared copies and alerts the team when it cannot reach them.
• A page blocked for more than half a second is noted in the app's log, and one frozen for 2 seconds or more is reported to the team, at most once every half hour from each browser.

Fixed

• On a phone, a long count no longer pushes a page's main button off the screen, a row's details no longer start or end a line with a stray "·", and a doc's project name is no longer squeezed to nothing.
• A preview deployment on devbench.sh talks to its own server again, instead of DevBench's, which signed its demo user straight out.
• The app is much faster. Up to six pages are kept open in the background so that going back to one keeps its place. Each of them used to redraw on every click anywhere: a project's Changes tab redrew every changed file, Chat every message, and Tasks every task, each time a page opened or the project refreshed. On an ordinary laptop, opening a project's tab took 2 to 3 seconds, and Tasks up to 8.
  • A page kept in the background no longer redraws. It keeps its scroll, drafts and open dialogs, and catches up once when you come back to it.
  • A refresh that changed nothing redraws nothing, and the task and file lists redraw only the rows that changed.
  • Measured on the same project with the processor slowed four times: opening Previews went from 5.0 s to 0.6 s, Docs from 3.7 s to 0.3 s, Projects from 3.7 s to 0.4 s, and Tasks from 7.8 s to 1.8 s.
• ⌘/Ctrl-click opens a new tab, everywhere. The top bar (Projects, Inbox, Docs, Automate, Ship, Insights), the row of open projects and a project's own tabs (Overview, Terminal, Chat and so on) are now real links. Hovering one shows where it goes, the browser's right-click menu offers "Open link in new tab", and ⌘/Ctrl-click or the middle button opens it in a new tab. Anything else that moves to another page, such as a row of the Apps table, a Settings tab or "58 of 69 tasks done", now opens in a new browser tab on ⌘/Ctrl-click instead of moving the page you are on. ⌘/Ctrl-click still selects tasks in a task list.

New

• Super admin. Whoever runs this DevBench has a new page, under the organization menu, listing every team and every person on it. From there a super admin can:
  • create a new team, which sends its first owner an invite link that works once, for 14 days;
  • invite someone into any team, as an owner, maintainer or member;
  • answer the website's access requests by giving the person a team of their own or a seat in an existing one;
  • switch a person's access off everywhere (or back on), and make other super admins.
• The page shows who is on the server and how many teams and projects there are. It never shows a team's projects or chats.
• The first super admin is set up on its own when this version is installed: the owners of the server's first team become super admins, unless the server already has one. manage.py superadmin grant <email> makes one by hand.

New

• Second Look. Each merge request in the Git tab has a 🔍 Second Look button. It runs an automatic review in the project's box with the project's own agent (Claude Code read-only, codex review, or OpenCode's plan agent): a separate pass over the merge request's diff, reporting specific problems (bugs, security issues, data loss, missing tests for risky changes), most severe first. The findings are posted as a comment on the PR/MR, shown on the button ("Second Look: 3 findings", "no findings"), and linked in the project's chat.
• Project Settings → Automation → Second Look on every merge request reviews each merge request the project opens, without anyone asking.

Fixed

• Found by Second Look's first review, of this release's own changes before they shipped:
  • Crew could keep a box "working" for hours on helpers that had died with their session. Its check for a running agent always matched itself.
  • Helpers the agent stopped itself now count as finished.
  • Reporting a finish line reached is limited to the project's agent or an owner. A stopped or replaced finish line can no longer be reached, or delivered late. Stopping one never starts a stopped box.
  • Gatekeepers' rules and enforcer live in a root-owned folder the agent cannot change. The agent is refused edits to its own hook settings. A protected file named from inside its own folder is still protected.
  • Only the box can report Gatekeeper events, and at most 30 a minute.

New

• Gatekeepers (project Settings → Gatekeepers): rules the box enforces on the agent before every action it takes, whatever the agent decides.
  • Tests before push: a git push runs your test command first, and is refused, with the failures shown to the agent, if the tests fail.
  • Protected files: edits to matching files or folders (*.env, migrations/, config/prod.yml) are refused, whether through the agent's edit tool or a shell command; reading them is fine.
  • Blocked commands: shell commands containing any of your phrases (git push --force, DROP TABLE) are refused.
  • Format after edits: your formatter runs on each file the agent edits.
  • Secret files (names only), on by default: .env, .env.*, *.env, *.pem, *.key, id_rsa*. The agent can't read or change them through any tool or shell command (cat, grep, source, cp, a python -c open(…)). When it tries, it is told only the variable names they hold, never the values. The app it runs still reads them.
• Saving applies them in the running box at once. Each refusal is listed under "Recently stopped". Only owners can change them. Every agent enforces them: Claude Code and Codex through their hooks, OpenCode through a plugin (not yet inside a Windows project's VM).