Repository navigation
Task workflows, session messaging, Playwright connector and browser drawer - #19
Merged
Merged
Conversation
Settings > Agents > Limits gets two rows: Workflow size (off, or Claude Code's workflowSizeGuideline small/medium/large/unrestricted, default medium) and Workflow agents at once (CLAUDE_CODE_WORKFLOW_MAX_CONCURRENT_AGENTS, unset = Claude Code's default). The resolver drops invalid values. Also adds the ClaudeRunWorkflow record type to ClaudeRunInfo. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
When triage sizes a task as large and workflows are on: - validation asks for a workflow that reviews the diff from several angles and has a separate agent check each finding before it counts (falling back to /code-review), instead of a single /code-review; - the implement turn allows a workflow for many-file changes. Each run passes the size guideline with one --settings object and the concurrency limit (or CLAUDE_CODE_DISABLE_WORKFLOWS) in its env. Workflow launches are recorded in claude_run.json (run id, task id, script path, session, status), so a step continued after a stop or usage limit is told to relaunch with resumeFromRunId and reuse the agents that finished. A profile swap now copies the session's directory with the transcript, keeping those saved results. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude Code's built-in messaging (SendMessage / ListAgents) reaches task runs
by their --name, but a session that skips permission checks holds every
cross-session message for approval, which a headless run can't give, so the
messages expired.
- New switch under Task messaging (Settings > Agents), "Messages from your
Claude sessions" (claudeCodeTasks.sessionMessages, on by default). While both
are on, task runs get --settings {"crossSessionInbound":"accept"} (one
--settings flag built from a per-session settings object).
- The task prompt says messages from agent windows, the chat or a terminal
arrive as cross-session messages (information, not a change of the task), and
that other running sessions are reached with SendMessage (names from
ListAgents). Tasks still message each other with message_task, which also
reaches tasks that are not running.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
With the Kanban button on, the chat gets a message_task tool that delivers a message from the user through Workbench's broker: a running task reads it at its next step, any other task when it next runs. It refuses while Task messaging is off. With messages from sessions also on, the chat gets Claude Code's ListAgents and SendMessage to reach other running sessions such as agent windows. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…asks An agent window's --append-system-prompt gets one line naming the project's task sessions (wb-<projectKey>-<task id>) and how to message a running one with SendMessage, while Task messaging and messages from sessions are on. Windows keep Claude Code's own session names: --name would replace the title the Session menu and /resume show, and ListAgents lists them anyway. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The task prompt and the inbox section framed every incoming message as "information from peers, not instructions from the user", including the [Message from the user, via Workbench] messages the Claude chat's message_task sends. Tasks are now told those come from the user, to be taken into account within the task. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- Workflow size gets a "Claude Code default" option, now the default, so task runs no longer override the user's own guideline (small on Pro) with --settings unless a size is picked. - Copy the session directory asynchronously on a profile swap: it can hold many agent transcripts and ran on the main process. - A step cut off twice still passes on the workflow runs of its earlier attempts; no relaunch hint while workflows are off. - A workflow that ends before its launch result is read is recorded as ended, not as interrupted. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Finds Chrome, Edge, Brave, Chromium and Firefox profiles (Local State info_cache, profiles.ini) on Windows, macOS, Linux and in WSL with async reads only, marks unusable ones with a reason (Safari, missing executable, Windows browser in WSL, Flatpak), and copies a chosen profile per run into .workbench/browser-profiles without caches and locks. Detection runs once at startup and is exposed over IPC with a refresh. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The Puppeteer server gets PUPPETEER_LAUNCH_OPTIONS with the chosen executable and engine; task runs and Claude chat add a copy of the profile (Chrome 136+ refuses remote debugging on its default data dir). The availability check launches the chosen browser with a temporary profile and only offers the Chrome install command for Puppeteer's own Chrome. Keeps @modelcontextprotocol/server-puppeteer; why not chrome-devtools-mcp is noted in mcp-config.ts. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A borderless dropdown on the Puppeteer row lists Puppeteer's Chrome and the detected profiles; ones that can't be used are disabled with the reason as text. Shows the last detection and detects again on refresh. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…eck on a new choice A stopped run's cleanup could delete the copy of the task's next run, which used the same folder: each run now gets its own slot, and older copies of the task are removed when a run starts. Choosing another browser checks the Puppeteer connector again, and a chat whose profile copy failed tries again on its next turn. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…sendmessage-listagents # Conflicts: # src/main/claude-code/__tests__/task-runner.test.ts # src/main/claude-code/settings.ts # src/shared/types/claude-code.ts
…the-puppeteer-connector # Conflicts: # src/main/insights/insights-executor.ts
…cker Two Firefox profiles named default (the regular install and the snap) showed the same label. Profiles from a snap or Flatpak install now read Firefox (Snap) · default. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Leftovers of removed installs (Firefox keeps a default-release-N profile per install) cluttered the picker. A profile is left out when none of the files a browser writes in use (prefs, cookies, history) changed in the last year. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
The picker showed a list cached in the window's storage until Refresh was pressed, so a restart kept an older list. It now takes the main process's in-memory startup result on open (no new scan). Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
…om scratch The picker no longer keeps the list in the window's storage (an older list survived restarts). Refresh also looks PATH commands up again instead of taking the cached answer. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
…profile ConnectorRow marks its control slot as a settings control, so the Puppeteer dropdown is the shared select with no custom styling. The refresh button is gone: the list is the one found when the app started. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
The browser picker lists Puppeteer's Firefox as a disabled item until it is in Puppeteer's cache, and offers a download button next to it (and next to Puppeteer's Chrome when its build is missing) that opens a terminal running npx puppeteer browsers install. The main process chooses the command, then watches the cache folder and sends the new list when the browser appears. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
…Puppeteer's browsers first Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
/snap/bin/firefox starts Firefox in the snap's sandbox, which can't open Puppeteer's profile in /tmp, so the launch failed with "Failed to launch the browser process". The snap's binary runs outside the sandbox. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…icker Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Runs copy the browser's profile, which an open browser keeps locked. The connector reads unavailable while the profile's lock names a live process, its switch's tooltip asks to close the browser, and runs leave it out. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…'t pose as the user Task runs skip permission checks, so accepting SendMessage from any local session let it drive an unattended run. Messages from your Claude sessions is now off by default, its description states the risk, and a run never accepts them when the user's own Claude Code settings refuse or hold them. Peer task messages can no longer fake the user's message marker. Revise and resync steps get the workflow relaunch hint, and message_task gets a label. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…aned up Copies are created 0700, leave out session restore state, Firefox's Sync account and nested caches, and remove a partial copy when copying fails. The chat makes a fresh copy per turn and removes it when the turn ends; copies older than a day that no live run uses are swept, and a stopped launch's sweep can no longer delete a newer launch's copy. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Firefox launches with Sync and Firefox Accounts off. Ubuntu's firefox wrapper is treated as the snap, and a snap that needs a newer glibc is marked unusable. Profile locks are read more carefully (host, browser process, both Firefox lock forms, network shares on Windows), and a half-unpacked Puppeteer Firefox isn't offered. The open-browser status no longer gets overwritten by a probe or kept for an old choice, and a stored browser choice follows its detected profile. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…n macOS Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The user's crossSessionInbound is read in any encoding Claude Code reads, and an unreadable file counts as refusing. Inbox entries carry a sender only Workbench sets, peer text is normalized and escaped so it can't pose as the user or Workbench, and the messaging switches are read again at each phase. Workflow relaunch hints survive phases recorded in between, and the agent-window hint is left out when runs won't accept messages. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Copies are swept at startup, removed on quit and when the choice no longer uses a profile. A stored choice only follows a profile of the same install, background Chromium and reused process ids are handled, Firefox ESR and per-user Windows Firefox are found, XDG_CONFIG_HOME is honored, and chmod is best-effort on filesystems without modes. Chat replies are saved without waiting for the copy's removal. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Runs log when they leave a connector out, the availability keeps the current browser's result and poll across quick switches, and only a missing download is shown as not downloaded. The download button runs the npx Workbench finds, Linux tries more terminals and reports when none opens, and the picker shows a download that couldn't start. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- A browser slow to answer the liveness check no longer counts as closed (which closed chat browsers for good); refused connections do - Connecting to a browser times out, so one that never answers can't hold up the bridge's queue - A screencast that failed to start is retried; a session whose endpoint changed drops its old link and pages first - Releasing the mouse outside the page still reaches it - A reopened drawer's first view isn't dropped before it subscribes - Closing a chat browser while it starts makes a reopen wait, so two browsers never share a chat's title - A background chat's last status no longer ends the New chat page's reply; the shown and background chats share one chunk handler - The registry poll doesn't probe browsers the bridge is connected to Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Workbench's temp folders (task runs, chat turns, chat browsers, the browser bridge, availability probes) carry the same owner file as profile copies. At startup a background sweep removes the ones no running Workbench or task process uses. A chat browser left by a crash is kept for its chat (when the chat still exists, for 7 days) with its profile copy and its open tabs; the next time that chat opens its browser with the same profile, its sign-ins and tabs come back. Everything else is removed, including the dead registry entries closed chat browsers used to leave. The availability probe removes its output folder only after the server exits, and its tests use a folder of their own. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
In WSL mode the browser bridge runs inside the distribution with its node, over stdin/stdout (a JSON header plus the frame's JPEG bytes, in native mode too), so it can reach the browsers' Unix sockets and their playwright-core. Workbench reads the distribution's registry through \\wsl.localhost and asks the bridge which browsers answer; workspace paths are translated before matching projects. Chat browsers run inside the distribution, with their folders in its ~/.cache/workbench/tmp, covered by the startup sweep and crash restore. Under NAT networking the chat's open-browser call goes through a Windows relay started over interop; with mirrored networking or WSL1 it calls Workbench directly. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Frames are JPEG quality 100 (lower ones blurred small text) and asked at the window's devicePixelRatio, following moves between screens. On HiDPI screens the browsers render that many pixels: the screens' largest scale factor (up to 2) goes into the runs' Playwright config as deviceScaleFactor, plus --force-device-scale-factor for Chromium, whose screencast otherwise stays at CSS size. Runs without a profile copy get it from a small config file in the app's data folder. Screenshots for agents stay in CSS pixels. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Only Claude chats' browsers belong to a project now: a task run's browser (working in the task's worktree or the project) matches none, so no window lists it, the drawer doesn't open for it and the bridge never connects to it. Task browsers stay headless and unseen, with Playwright's own viewport. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The probe's output folder goes a second after its server exits. Under fake timers vi.waitFor moves the clock 50ms a try, so that second takes about one of real time, right at waitFor's default timeout, which the Windows runner missed. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The drawer shows a chat's own browser, so it is mounted on the Claude page alone and opened from the chat's top-right button; the handle on the window's edge is gone from every page. Chat browsers keep running while the user is elsewhere. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The Claude page's top bar starts below the window drag strip, where the drawer starts: while the drawer is open its top line continues the drawer's top border to the chat history. The history toggle and the browser button share the first row, as far from that line as the button is from the right edge, with New chat below; the extra row the button needed under the window controls is gone. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Opening or widening the browser drawer until the chat is narrower than 500px closes the chat history sidebar; narrowing or closing the drawer opens it again once the chat keeps 500px with it. Only a history the drawer closed reopens: one the user closed, or toggled since, stays as the user left it. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Ctrl+V in an agent window pasted clipboard text only, so a screenshot or a file copied in a file manager pasted nothing. Pastes now go through the browser's paste event: copied files paste as their paths (Claude Code attaches image paths as images, in WSL as /mnt/c paths), and an image with no file behind it sends the key Claude Code reads clipboard images on (Alt+V on Windows, Ctrl+V on Linux). Under WSLg a file copied in Windows arrives as an empty paste, so the main process reads the Windows clipboard's files; a screenshot can arrive as text holding the bitmap's bytes, which is treated as an image. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The Store's Firefox keeps its profiles in its package's LocalCache and starts through an app execution alias, so the browser list never showed it. Detection now looks through %LOCALAPPDATA%\Packages for any package holding Firefox profiles and lists them as "Firefox (Microsoft Store)". Choices keep the alias, and runs launch the package executable it points at: Playwright rejects the alias itself, and the target's path changes with every Store update. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Firefox makes a placeholder `default` profile on its first run next to the `default-release` one it uses, and the picker listed both. A profile with none of the files a browser writes in use was never opened: it is now left out once its folder is a week old, while a profile just made in about:profiles is still listed. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
While tools run in WSL, Playwright runs inside the distribution, so Windows browsers were listed but disabled. A Windows Firefox profile is now copied into the distribution and run by a Firefox there, keeping its sign-ins (Firefox's cookies and saved passwords travel with the profile): an installed Firefox at least as new as the version that last wrote the profile, else Playwright's own Firefox with --allow-downgrade, since Firefox refuses a profile a newer version wrote and only the run's copy can be harmed. Without either, the picker offers to download Playwright's Firefox. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
While tools run in WSL, Playwright can't launch a Windows browser (it talks to browsers it launches over pipes, which interop doesn't carry) and WSL2 with NAT networking can't reach the Windows 127.0.0.1. The Playwright server's command for a Windows Chromium browser is now a launcher run by the distribution's node: it starts the browser through interop on a profile copy kept on Windows, relays a local port to its DevTools port through a relay run by Workbench's own executable as Node, and starts Playwright with --cdp-endpoint, so nothing is needed on Windows besides the browser. The relay's keeper closes the browser and removes a temporary profile when the launcher goes, even when it is killed. The browser starts with no first tab (Playwright's screencast gets no frames from a tab opened before it attached), without --force-device-scale-factor (it leaves the drawer's screencast blank once the viewport is set) and with Edge's Alt+Tab tabs off (they stay there when the browser is killed). These profiles are marked "Saved sign-ins don't carry over": Chrome and Edge encrypt cookies with an app-bound key that only opens in the browser's own profile folder. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A checkout can keep a release tag made before main's history was rewritten: `git fetch` never moves a tag that already exists locally. Such a `v<major>.0.0` tag shares no history with main, so counting the pull requests merged since it took every one main ever merged (1.11.x instead of 1.5.x). The tag must now be on main's history, else the version is kept and the message says how to take the remote's tags. `--dry-run` prints the version without writing it. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The pages of the Claude chats' browsers rendered light whatever theme Workbench showed (Playwright emulates a light color scheme by default). The drawer now sends Workbench's theme, the `dark` class on the window, also when it changes, and the browser bridge gives every page it tracks that theme as its `prefers-color-scheme` (page.emulateMedia), new tabs included, so sites render dark or light like the rest of the app. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…lay scales On a 1.25x or 1.5x display, a Chromium chat browser's drawer stayed blank once a page loaded: with a fractional --force-device-scale-factor Chromium sends no screencast frames after the drawer sets the page's viewport. The switch is now rounded up to a whole factor while pages keep rendering at the real deviceScaleFactor, so frames come at the device-pixel size the drawer asks for and stay sharp. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A window's saved output is cut wherever it was last saved, which can be inside an OSC 8 link: Claude Code redraws its status line's PR link all the time. Replayed on the next start, the link stayed open and the resumed session never closes a link it didn't open, so xterm drew a dashed underline under everything after it. The replay now ends with a link close and an SGR reset. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The drawer took a tab's icon from Playwright's registry entry, which reads as Playwright's own build both for a Windows Chromium browser attached over CDP (no executable recorded) and for a Windows Firefox profile run by Playwright's Firefox in WSL. A chat browser's tabs now show the browser the chat chose, which its holder reports. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Agent windows paste from the browser's paste event now (copied files and images included), not navigator.clipboard.readText(). The integration tests fire that event: its text goes to xterm.paste(), an empty paste on Linux asks for the clipboard's files and then sends the empty paste, and a path that can't be translated is pasted as it was. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Widening the drawer until the chat is narrower than 500px closes the chat history, then collapses the app's sidebar to its icon rail (not saved), and the drawer now widens until the chat keeps 500px beside the collapsed sidebar instead of stopping at half the window. Narrowing or closing the drawer brings them back in the other order; only the drawer growing closes them and only its shrinking reopens them. A sidebar the drawer closed that the user opens again narrows the drawer by the room it takes, so the chat keeps its width. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… browsers - A stopped bridge (quitting) is never started again by a probe, and a spawn still starting when it stops aborts and removes its folder; node is looked up before the temp folder is made - A restorable chat browser keeps its profile copy until its holder has started, and a fresh copy made for a reuse goes when the start fails - The holder's start has one deadline again; restoring tabs stops once the holder exits instead of matching its error text - Frames are joined once whole, not copied again for every chunk, and the newest frame of a page waiting for a busy pipe is sent on drain - A bridge that starts again (WSL probes) gets the theme too - One registryOs(); misplaced doc comments put back Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…r slow browsers - Chat browsers follow Workbench's theme on every page (the window's useBrowserDrawerSync sends it), not only while the drawer is shown - In WSL mode a liveness check never starts the bridge: it starts once a drawer opens, and until then the chat browsers Workbench holds open count as live - The keeper of a Windows browser run from WSL starts with the browser and reads its port from DevToolsActivePort, so a browser slower than the launcher's wait is still closed and its temporary profile removed - One pixel scale rule (pixel-scale.ts) for the browsers' density and the drawer's frames Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
parseFileUris reads file:// URIs from a Linux clipboard, but Node's fileURLToPath refuses a path without a drive letter on Windows, so its tests found nothing there. The URIs are decoded as POSIX paths instead (files on other hosts are still left out). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
At a fractional device pixel ratio (1.25, 1.5) the drawer's page area usually starts between two device pixels, so the display resampled every frame and blurred all of the page's text. The page's viewport is now a whole, even number of device pixels, and the canvas is moved onto the pixel grid, also when the window resizes. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The drawer froze and dropped clicks, most of all with several tabs open: the browser sent frames as fast as the page changed, whether or not the window had drawn the last one, and every mouse move was replayed in order, which a busy page answers late (a click came 4.2s after it was sent in a test). Frames now go one at a time, the window acknowledges each one it draws, and frames waiting meanwhile are replaced by newer ones. Input has a queue per page, with mouse moves and wheel steps waiting behind a slow one merged into the newest; the same test now takes 0.2s. No viewport change, screencast call or input step can hold up the ones after it for longer than 5 seconds. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
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.
Summary
Task workflows
/code-review), and the implement turn may use one for many-file changes. A change request keeps the task's own size for whole-task validation; the follow-up's triage only sets the revise turn's model and effort.workflowSizeGuidelinewith--settings) and Workflow agents at once (CLAUDE_CODE_WORKFLOW_MAX_CONCURRENT_AGENTS). With triage set to Fixed, tasks never ask for a workflow; the size still applies to ones Claude Code starts itself.claude_run.jsonwith the commit they started on. A step continued after a stop or usage limit is told to relaunch withresumeFromRunId, so finished agents reuse their results; a review is not relaunched once the code changed (a finished revise, implement or resync, or a different HEAD). A profile swap copies the session's folder asynchronously, keeping those results.Parallel tasks
message_taskwhen both need the same lines (git merges changes to different parts of a file by itself).git merge-tree --write-tree) and tells both tasks the conflicting files and lines. Without a usable git it falls back to the shared-file notice.Messaging between tasks and sessions
claudein a terminal can message running tasks with Claude Code's SendMessage (task runs getcrossSessionInbound: accept). Its description says plainly that task runs skip permission checks. A run never accepts when the user's own Claude Code settings refuse or hold such messages (read in any encoding Claude Code reads; an unreadable file counts as refusing).<message sender="user|task|workbench">), and only the tags that delimit them are escaped in message text, so a peer task can't pass its text off as the user's note or a Workbench notice. Messages the user sends from the chat are taken as the user's notes.Claude chat
message_task: a running task reads the message at its next step, any other task when it next runs. With messages from sessions also on, it gets ListAgents and SendMessage to reach other running sessions such as agent windows. While Task messaging is off, the prompt points toupdate_taskandrequest_changesinstead.Agent windows and Live agent
--append-system-promptnames the project's task sessions (wb-<projectKey>-<task id>) and how to message a running one. It is left out when the user's own settings keep runs from accepting.Task view
Playwright connector (replaces Puppeteer)
@playwright/mcp(pinned 0.0.83,PLAYWRIGHT_MCP_PACKAGE): accessibility-snapshot tools with auto-waiting, Chromium, Firefox and WebKit. The default is Playwright's own Chromium; Playwright's Firefox and WebKit are offered as downloads (npx -y @playwright/mcp@<version> install-browser <name>in a terminal; the picker shows the download and the connector is checked again once it finishes). Data saved for Puppeteer (connector id, switch, browser choice in settings, tasks, drafts and chats) is renamed where it is read; a custom server with a built-in id moves to<id>-custom.XDG_CONFIG_HOME), with unusable ones disabled and the reason shown (Safari, missing executable, Windows browser in WSL, Flatpak, a snap that needs a newer glibc). Profiles nothing used for a year are hidden, and ones a browser never opened (Firefox's placeholderdefault) are left out after a week; snap and Flatpak installs are told apart. Firefox from the Microsoft Store is found by its package's profiles and run through its app alias's target. Chromium-family profiles run their executable on a copy with the sandbox on and without the switches that keep copied sign-ins from decrypting; the installed Firefox runs over WebDriver BiDi (moz-firefox) with Sync and Firefox Accounts off; snap Firefox through the snap's own binary.--allow-downgrade), keeping its sign-ins. Windows Chrome and Edge run on Windows through a launcher (interop) and a DevTools relay run by Workbench's executable, with Playwright on--cdp-endpoint; a keeper started with the browser closes it and removes its temporary profile when the launcher goes, even after a slow start or a kill. Their sign-ins don't carry over (Chrome and Edge's app-bound cookie encryption).about:blank, headless, into a temp output folder). Runs pass--output-dirto their temp folder, so Playwright's snapshots and screenshots don't land in the worktree.browser_run_code_unsafe, so the chat still can't change the project's files.Browser drawer
--headlesson every server). Claude chats' browsers show in a drawer on the Claude page, in Workbench's own style: a tab per page (with the chosen browser's icon, and a + for new tabs), back, forward, reload and an address field, and the live page sized to the drawer and usable with the mouse and keyboard (including paste). Its pages follow Workbench's light/dark theme. Closed, the chat reopens it from a button in its top-right corner; other pages have no drawer. Task runs' browsers stay headless and unseen: no drawer lists them and nothing connects to them.ms-playwright/b) and connects to the bound browsers with each one's own playwright-core from the npx cache, so nothing of Playwright ships with the app. It sizes the viewed page's viewport to the drawer, streams JPEG frames (quality 100) to that window only and replays input; a frame is joined once whole, and a page's newest frame waiting on a busy pipe is sent when it drains. Frames are asked at the window's devicePixelRatio, and on HiDPI screens the browsers render that many pixels (deviceScaleFactor, plus--force-device-scale-factorfor Chromium, rounded up to a whole factor since a fractional one stops its frames); screenshots for agents stay in CSS pixels.--endpointand kept open between turns and while switching chats or pages. The user can sign in or get past MFA by hand, then let the chat continue. The chat can't close the whole browser; closing its last tab, deleting the chat or quitting does. Each chat remembers its drawer state.\\wsl.localhost, and chat browsers run in the distribution (folders in its~/.cache/workbench/tmp). Under NAT networking the chat's open-browser call goes through a Windows relay started over interop. A liveness check never starts the bridge: it starts once a drawer opens, and until then the chat browsers Workbench holds open count as live.Terminals
osascript(it used to block the main process, also while the first-use Automation prompt waited). kitty, wezterm and alacritty fall back to Terminal.app when they don't start.Developer
--dry-runprints the version.🤖 Generated with Claude Code