Repository navigation
fix(computer): install only the Chromium build the computer launches - #829
Merged
dustywusty merged 2 commits intoOct 27, 2026
Merged
dustywusty merged 2 commits into
dustywusty merged 2 commits into
Conversation
`playwright install chromium` on 1.62 installs two browsers: the full
Chromium and the separate chromium-headless-shell. The computer only
ever launches with a named channel ("chromium", or "chrome" for a local
browser), and Playwright's getExecutableName returns the full build for
any Chromium-alias channel, headless or not. The shell is used only by a
bare chromium.launch(), which shipped code never does.
The shell was 340 MB unpacked and about 106 MB of the agent-computer
image's 426 MB browser layer. The bunx package cache Bun leaves under
/usr/local/install/cache and /root/.bun is another 35 MB unpacked.
Install with --no-shell and remove both caches, in the agent-computer
image and the all-in-one image. The real-browser CI job now installs
the same way, so its headless run fails if anything needs the shell.
The one test that launched bare (sign-in.test.ts) now names the channel,
as the computer does.
dustywusty
requested review from
MikeRyanDev,
davidmckayv,
guidovizoso,
mxmzb and
tylerslaton
as code owners
October 23, 2026 10:26
Codex Review SummaryThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
guidovizoso
approved these changes
Oct 27, 2026
guidovizoso
left a comment
guidovizoso
left a comment
Collaborator
There was a problem hiding this comment.
Smaller first download and smaller upgrades, with nothing for the person to do. Good one.
9 tasks done
davidmckayv
added a commit
that referenced
this pull request
Oct 28, 2026
* Let each person pick a skin for OpenBot in Settings Adds app/src/skins: a registry of skins, a per-browser choice stored like the dark theme and applied before first paint as data-skin, a Skin row in Preferences, and /s/$screen for screens only a skin has. A skin may replace the frame around the app's pages and scope its tokens to its own data-skin; tests hold every skin's stylesheet to its own scope and to defining dark for every light token. OpenDuet, OpenDots and OpenMuse are registered as stubs. * Give OpenDots a real skin: its rail, Dots sidebar, top bar and computer pane around OpenBot's pages The OpenDots stub becomes OpenDots' own frame drawn over OpenBot's data. A 48px rail (home, sidebar toggle, new chat, Bots, activity, skills, connections, admin, settings, account) sits beside a 220px sidebar whose DOTS are the Bots from the agent list, each wearing one of OpenDots' four plush avatars chosen by its id, and whose RECENT CHATS are OpenBot's conversations with search, pin and delete. A Dot reopens its latest chat or starts one; its "..." opens the Bot's page. The top bar carries OpenDots' breadcrumbs, a pause and resume for the Bot on screen, and the computer toggle: on a conversation it drives the conversation's own computer pane, elsewhere it opens OpenBot's computer panel in OpenDots' right pane. At phone width the sidebar becomes OpenDots' hamburger drawer. Every page inside is OpenBot's own, chat included, in OpenDots' tokens for light and dark (the dark set derived with OpenDots' own lightness flip). Scheduled & activity and Memories are skin screens that put OpenBot's routines and memory lists under OpenDots' page heading. Spaces, pages, the Brief and the pause-everything switch are left out: OpenBot has nothing behind them. Tests draw the frame against OpenBot's own agent routes and hold it to: every OpenBot surface reachable, Settings reachable on every page even when nothing loads, DOTS listing the agent list, a Dot opening a chat with its Bot, the computer toggle driving the conversation's pane, and per-Bot pause reaching the server. * Show OpenDots' coworker home on a new conversation with a chosen Bot OpenDots opens a Dot on its persona: the plush, the name, what it does and a way to edit it, above an empty conversation. OpenBot's new-conversation page leaves that space empty, so the frame draws the persona there from the Bot's profile and hides it as soon as the page shows a message. Also keeps the composer and its notices off the pane's edges where the column is no wider than the composer, drops the rule under a conversation's header so it reads as one band with the top bar, and gives the disabled send button OpenDots' pale disc. * Dress a conversation's own computer pane as OpenDots' right pane The top bar's computer toggle opens the conversation's own Computer pane, so it now wears the same grey ground, ruled header and selected-tab title as the frame's pane elsewhere. * Add the OpenDuet skin: a phone-width frame and a Tasks screen on OpenBot's data OpenDuet's look, rebuilt from OpenBot's primitives, Tabler icons and OpenDuet's tokens, so a person can pick it in Settings and keep everything OpenBot does. The frame is one column at most 430px wide. On Chat and Tasks a header floats at the top with the menu disc, OpenDuet's mic orb and the person's initial (their Settings), and a frosted tab bar floats at the foot. The orb presses the composer's own dictation rather than running a second one, so recording, transcription and where the words land stay OpenBot's; on a screen with no composer it opens Chat and starts there. Every other page opens over the app with a back disc and its own heading in a title pill. The side menu lists Chat, every conversation and group (pinned first, pin and delete on right-click), New conversation, and under Manage: Bots, Connections, Skills, Memory, Routines, Responsibilities, Approvals, Notifications and, for administrators, Admin. Settings is never folded away. The frame keeps OpenBot's channel socket and the browser notification the sidebar's Bots badge sent. Tasks, at /s/tasks, shows only what OpenBot keeps: pending approvals and Bots' questions, decided and answered on their cards with OpenBot's own mutations; active responsibilities and switched-on routines; and, under Up next, background research's suggestions with Start and Hide. Left out because OpenBot has no backend for them: Make, the morning brief, triage, the setup checklist, first-day onboarding, reply chips, channel icons and colours, Archived, the "Match phone" theme and adding tasks. * Let the OpenDuet orb wake only when the composer starts or stops recording The orb watched the whole conversation for the composer's recording controls and set its state on every change to the page, which during a streamed reply is every word. It now sets state only when recording actually starts or stops. * Set OpenDots' composer on the canvas colour, as OpenDots draws it in dark OpenBot's composer sits on the card colour, which in this skin's dark set is a step above the canvas; OpenDots' composer is the canvas with a hairline around it. * Fix what OpenDuet's live pass found: the orb, the approval card's glyph, the composer's edge Driving the skin against a real OpenBot found three things. The orb rested while a Bot browsed with its computer: it read only the roster's busy flag, which arrives over the socket and missed the person's own run. It now also reads the composer's Stop control, which OpenBot draws for exactly as long as a turn runs, and waits 700ms before thinking so the moment Stop shows while a conversation opens is not a flash of "Thinking" on every open. A waiting computer_run_command was drawn with the browser's globe, because the card asked whether the tool was the computer's before asking whether it was a command. The kind is now worked out once, commands and files first. A conversation's composer ran flush to the 430px column's edges with its focus ring on the screen's edge. The skin insets OpenBot's composer wrapper by 12px, as OpenDuet floats its own. * Build the OpenMuse skin: one agent, a phone-shaped column, five tabs OpenMuse is a personal agent shaped like a phone. This rebuilds it on OpenBot's own components, Tabler icons and tokens so a person can pick it in Settings and keep everything OpenBot does. The frame is a single 760px column: a menu button, the centre Bot as OpenMuse's capybara with its name and a live status line, a bell for the approvals and questions waiting on the person, and a floating pill of Chat, Activity, Ideas, Goals and Apps. The centre Bot follows the conversation on screen, then the person's choice, then OpenBot's default Bot; a sheet behind the name switches it. Chat is OpenBot's own chat route under the frame, restyled through the skin's tokens and scoped CSS, so the computer, approval and component tools keep working. Activity, Ideas, Goals and Apps read OpenBot's Bot activity, approval inbox, suggestions, routines, responsibilities, connections and memory through the existing query and mutation factories. The menu reaches every OpenBot surface, Settings and Admin included. Mail, calendar, page watches, PDF filling and task plans have no OpenBot backend and are left out. Light is OpenMuse's palette; dark is a night version in its spirit. The tests render the frame against OpenBot's API shapes and check that every surface is reachable and each tab reads and writes the right data. * Clear a skin this build no longer ships from the page on load index.html applies the stored skin before first paint without knowing which skins exist, so a removed skin's name stayed on <html> for the whole visit while the app drew the default. The store now settles the attribute when it loads. * Note the skins in the changelog * Give the OpenDuet skin OpenBot's identity: the deployment's name, OpenBot's orb and avatars A skin keeps OpenDuet's look and takes its identity from the OpenBot deployment. The menu header and the orb's "Talk to …" now read the product name from brand.yaml (appConfig.brand.productName), as OpenBot's own sidebar does, and nothing a person reads says OpenDuet; the skin's picker name stays OpenDuet, since that names the look. The orb's face is OpenBot's AgentOrb, as sign-in and onboarding draw it, in place of OpenDuet's mic-orb robot, which is removed. Its behaviour is unchanged: it presses the composer's dictation, listens while recording or typing, and thinks while the Bot works. Conversation avatars are each Bot's own, drawn from its avatarSeed in OpenBot's ChannelAvatar, and Tasks uses OpenBot's words for decisions: Allow once, Always allow here, Deny, Start as task, Dismiss. * Wear the deployment's brand in the OpenDots skin, keeping only OpenDots' look A skin changes how OpenBot looks, never whose product it is. The wordmark and the home links now name the product from brand.yaml (appConfig.brand.productName), the rail's monogram is OpenBot's AgentOrb, and every Bot row and the coworker home draw OpenBot's own avatar from the Bot's avatarSeed. The plush avatars and their images are gone. The words follow OpenBot: BOTS, "Bots" in the breadcrumb, "Edit Bot". The rail, sidebar layout, top bar, computer pane, density, type and palette stay OpenDots'. The skin's own name in the picker is still OpenDots, since that names the look. New tests render the frame and hold it to the product name, to never showing "OpenDots" or "Dots", and to drawing each Bot with OpenBot's avatar from its seed; both fail on the previous skin. * Wear the deployment's brand in the OpenMuse skin Skins keep their look but take their identity from the OpenBot deployment. The OpenMuse skin now names itself with the brand.yaml product name, the same one OpenBot's sidebar shows, and never says "OpenMuse" to a person; only the picker still calls the look OpenMuse. The capybara is gone. The top bar, the Bot switcher, the Your Bot card, approval cards and conversation rows draw each Bot with its own OpenBot avatar, and the deployment's AgentOrb stands in when there is no Bot yet. Copy uses OpenBot's words: Bots, Connections, Approvals. A regression test renders the frame on every tab and checks that it shows the product name, never shows "OpenMuse" in text, labels or alt text, and draws no capybara. * Read the Anthropic key from either header in the AG2 test Anthropic's SDK 2.3.0, which CI now installs for the AG2 harness, sends the key as Authorization: Bearer instead of x-api-key. The agent's key still reaches Anthropic; the test only looked for the old header, so it failed on every pull request. It now accepts either, and passes with Anthropic's SDK 1.17.0 and 2.3.0 alike. (cherry picked from commit a1391e7) * Build OpenDuet's Tasks from OpenBot's own lists, and give pages OpenBot's width From Guido's review of the skins. Tasks re-implemented OpenBot's approvals, suggestions and routines, and its copies had drifted: its own tool-name formatter, its own decision buttons, approval cards that were all regions with the same name. Tasks now draws OpenBot's components unchanged (WaitingForYou for approvals and questions, BotResponsibilities, RoutinesList and SuggestionsInbox) under OpenDuet's Active and Up next switch and section headings. What an action will send, every choice and every confirmation are OpenBot's. The skin keeps only the badge's count, and a test renders WaitingForYou and holds the badge to the rows it draws. The frame no longer writes an attribute onto a page's own heading. A page opened over the app has a back disc and keeps its heading as drawn. Only Chat and Tasks are phone-shaped now. Every other page keeps OpenBot's width on a wide screen, and a conversation with its computer or its Bot's details open grows by OpenBot's 400px pane, so the computer sits beside the conversation as it does in OpenBot. Every animated control in the frame says what reduced motion does, and a test holds new ones to it; the orb is AgentOrb, which reads the same preference. The skin's description no longer calls a Bot an agent. * Make the OpenDots skin pass AA contrast, still for reduced motion, and reuse OpenBot's parts Review on #856 found the skin's page label and description below WCAG AA, house-style copy problems, no reduced-motion handling, and pieces rebuilt that OpenBot already has. - Contrast: every text colour now reaches 4.5:1 and every icon 3:1 on the surface it sits on, in light and dark. Greys that fell short are darkened along their own hue; the YOUR WORKSPACE label grows from 8px to 10px. - Copy: the picker description no longer says "coworkers"; the schedule screen is Routines, as OpenBot names it, at /s/routines, without "run on the server"; empty states and errors say what will appear and what to do next. - Reduced motion: the frame's chrome and the menus, tooltips, sheets and dialogs it opens stop transitioning and animating under prefers-reduced-motion. - Reuse: the phone drawer is OpenBot's Sheet (focus trap, Escape, scrim) at OpenBot's own 768px breakpoint, and older conversations load through OpenBot's ChannelPagination. New tests measure every colour pair against WCAG in both themes, hold the reduced-motion rule and the chrome markers in place, and open the phone drawer to reach every surface. * Save the skin on the account, share one navigation list, and give skins data-slot hooks The skin was a per-browser choice, so the next person on a shared machine inherited it. It is now a user preference (`skin`, default "openbot") that the server validates against SKIN_IDS, which move to shared/. The picker draws a skin at once and saves it through an optimistic preferences mutation that puts the skin back and says so when the save fails. SkinAccountSync, mounted in _authed, applies the account's skin whenever the preferences load or change, so the account wins over the browser's cache. localStorage stays only as the pre-paint cache, and the central sign-out clears it. The picker names the default skin after brand.yaml's product name and shows each skin's description. OpenBot's destinations (Bots, Capabilities, Connections, Admin for admins, Settings) live in lib/navigation.ts. The sidebar renders its footer and user menu from that list, so skins can read the same one. Skins styled OpenBot through Tailwind classes and ARIA copy. The composer, its input and buttons, the conversation header, the composer column, the detail pane and the page title now carry data-slot hooks for them to use. The Connections listing rules move from the route file to lib/plugins/listed.ts, so OpenMuse's Apps screen stops importing a route. * Build the OpenMuse skin's tabs from OpenBot's own components The OpenMuse skin rebuilt approvals, routines, responsibilities, suggestions, memory and connections, and its copies had already drifted from OpenBot: a different tool-name formatter, "What it will send" hidden behind a toggle, and its own routine delete flow. A new approval choice or routine field would have been missing from the skin with nothing failing. Each tab now puts OpenBot's own sections for the centre Bot under OpenMuse's titles. The bell and Activity use WaitingForYou and BotNeedsYou; Activity adds the Bot's activity sections; Ideas uses SuggestionsInbox and ProactiveResearchSettings; Goals uses RoutinesList and BotResponsibilities, with OpenBot's Delete dialog. Apps is OpenBot's Connections page. The skin's copies and its formatter are gone. A tab is lit only on its own screen, not on every other page. Pages that are not a tab, and a conversation with its computer or settings open, take OpenBot's own width on wide screens; the top bar and tab bar stay centred. The chosen Bot is remembered per signed-in person, so the next person on a shared browser starts from OpenBot's default Bot. Buttons, tabs and sheets respect reduced motion, and copy follows OpenBot's style. * Draw OpenDuet's menu from OpenBot's navigation list and dress pages by their hooks The menu's destinations are OpenBot's one navigation list now (useNavigationItems: Bots, Capabilities, Connections, Admin for admins, Settings), so a destination added to OpenBot reaches this menu too. Chat, the conversations and Tasks stay the skin's own. The skin's copy of the list is gone, and the avatar disc takes Settings from the same list. Every selector that reached OpenBot's markup by its classes or its words now uses the data-slot hooks: the composer column for the inset, the composer's stop control for "working", its input for the caret. Only the three dictation controls, which have no hook, are still found by the names OpenBot gives them, and a test holds those to the real composer. A page opened over the app has its title back in the pill, read from its page-title hook and never written to; the title stays for screen readers. Where the page draws its own back link, the chrome's back disc goes where that link goes and the page's own bar is hidden, so there is one way back. * Give the composer's dictation buttons their own hooks A skin that restyles or drives dictation found the three buttons by their aria-labels, so rewording one would break it without a test failing. They carry composer-dictate, composer-dictate-cancel and composer-dictate-stop. * Build OpenDots' destinations from OpenBot's navigation list and style pages through their hooks The rail and sidebar now draw OpenBot's destinations from useNavigationItems(), the same list OpenBot's own sidebar renders, so a destination added there reaches this skin too. The skin keeps only what is its own: BOTS, RECENT CHATS, and its Routines and Memories screens. Bots sits under the list of Bots; Capabilities and Connections are in the rail and the phone drawer; Admin and Settings close the sidebar. OpenBot's Bots badge, which also sends the browser notification, is mounted once: on the rail's Bots button, or on the menu button on a phone. skin.css no longer names OpenBot's Tailwind classes. The composer, send button, conversation header, composer column and detail pane are styled through the data-slot hooks those pages publish for skins. Tests hold the rail and sidebar to navigationFor() for a member and an admin, and fail any selector in skin.css that names a class other than .dark. * Find the composer's dictation controls by their hooks in the OpenDuet orb The orb found the composer's dictate, cancel and stop controls by their accessible names, so rewording one for a screen reader would have broken it silently. It uses the composer-dictate, composer-dictate-cancel and composer-dictate-stop hooks now, and a test fails if any selector in the skin, in its stylesheet or its code, targets an aria-label. * Hook the OpenMuse skin onto OpenBot's navigation list and data-slots The menu sheet kept its own list of OpenBot destinations, which would miss any page OpenBot adds. Its destinations now come from the shared navigation list OpenBot's sidebar draws, filtered for the signed-in person; conversations and the skin's tabs stay as they were. The stylesheet reached OpenBot's components through Tailwind classes and ARIA labels, so restyling the composer or the conversation header in OpenBot would have broken the skin without a test failing. It now uses only the data-slot hooks those components publish: composer, composer-input, composer-attach/send/stop, conversation-header and page-title. The system font is set through the font-sans token rather than by overriding OpenBot's font classes. * Say in the changelog that the skin follows the account * Read the Anthropic key and API version the 2.x SDK sends in the LangGraph AG-UI test anthropic 2.x sends the key as Authorization: Bearer instead of x-api-key and dates its Messages API version later than 2023-06-01. The provider-boundary test asserted both by their 1.x values, so it failed once the unpinned requirements resolved 2.x. It now reads the key from either header and checks the version is a date; the endpoint, key, model and body it asserts are unchanged. (cherry picked from commit 99685ee) * Launch the Chromium build CI installs in the enter-text test CI installs Chromium with --no-shell since #829, and enter-text launched Playwright's default, the separate headless shell, so it failed to start. It now launches channel "chromium", the build the computer itself launches. (cherry picked from commit f220198) * Record isBrokered's new home in the duplicate-definition guard The Connections page's helpers moved from the route file into lib/plugins/listed.ts, and the guard's list of known brokered-provenance sites still named the route file. * Wait for home's composer, not its greeting, when a skin switch lands home Home's heading became a personal greeting in #861, so the test that a skin switch on a skin-only screen lands on a drawn home waited for wording that no longer exists. It now waits for the home composer.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What this changes
playwright install chromiumon Playwright 1.62 installs two browsers: the full Chromium and the separatechromium-headless-shell. The computer never starts the shell. This PR stops installing it, and removes the bunx package cache that is left in the same layer.What it saves
openbot-agent-computeragent-computer is the biggest image a desktop install downloads. On Apple silicon a first install downloads about 1.24 GB of OpenBot images, and this takes that to about 1.13 GB (−9%). Each upgrade that changes this layer downloads 114 MB less too. The all-in-one
openbotimage gets the same change.How this was measured. I took the browser layer of the published v0.2.0 image and unpacked it. I packed it again unchanged, and again without the files this PR removes (
ms-playwright/chromium_headless_shell-1234,/usr/local/install/cache,/root/.bun). Both were compressed with zstd level 3, the same as BuildKit's default. To check the method, I compared the unchanged repack with the real layer: arm64 gave 426.0 MB against the published 426.6 MB, and amd64 gave 401.1 MB against 408.0 MB. "Saved" is the difference between the two repacks. The release build will give the exact sizes.What the arm64 layer held (1.38 GB unpacked):
ms-playwright/chromium-1234ms-playwright/chromium_headless_shell-1234--with-deps)Why the shell is never used. The only launch site,
agent-computer/src/profiles.ts:476, always passeschannel: BROWSER_RUNTIME.channel, which is"chromium"or"chrome"(browser-runtime.ts:45). In playwright-core 1.62.1,getExecutableNamereturns"chromium"(the full build) for any Chromium-alias channel, headless or headed. It returnschromium-headless-shellonly when no channel is given. The one other launch in shipped code,scripts/self-hosted-learning.ts, is headed and passes a channel.Changes:
agent-computer/Dockerfileand the all-in-oneDockerfile:install --with-deps --no-shell chromium, andrm -rf /root/.bun /usr/local/install/cachein the sameRUN..github/workflows/ci.yml(computer-browser): installs with--no-shellas well. That job says it uses "the one the computer image installs", and now it does, so its headless run fails if anything needs the shell.agent-computer/tests/sign-in.test.ts: this was the only barechromium.launch(). It now passeschannel: "chromium", the way the computer launches. Without that it fails against a browser install that has no shell, which is how I found it.tests/compose.test.tsanddocs/deployment.md: updated for the new install line.Verification
PLAYWRIGHT_BROWSERS_PATHset to an empty directory,bunx playwright install --no-shell chromiuminstalls onlychromium-1234andffmpeg-1011.sign-in,session-cookies,follows-popup,demonstration-live,webrtc-proxy,web-bot-auth-browser,handoff-page-restore, withCOMPUTER_BROWSER_MODE=headless) gives 21 pass, 0 fail. Before the test fix,sign-in.test.tsfailed withExecutable doesn't exist at …/chromium_headless_shell-1234/….web-bot-auth-browser.test.tsheaded: 1 pass.bun test tests/compose.test.ts: 17 pass.headed-browser.test.ts. CI runs both.Where it runs
Boundary and audit
Changelog
Unreleased.