Skip to content

Clawd mode and session messaging for Claude chat, a fuller browser drawer, accounts and window improvements - #20

Merged
ArtyomVancyan merged 74 commits into
mainfrom
develop
Nov 14, 2026
Merged

ArtyomVancyan merged 74 commits into
mainfrom
develop

Conversation

@ArtyomVancyan

@ArtyomVancyan ArtyomVancyan commented Nov 7, 2026 •

Copy link
Copy Markdown
Member

Claude chat: the Clawd button

  • Clawd button (between GitLab and Kanban): a chat's turns get what an agent window has: every tool (changing files, any command), the user's own MCP servers, and the agent windows' permissions (skipped when Settings skips them, else auto mode). Turning it off keeps the same Claude Code session; later turns are read-only again. Tools only a Clawd turn can call show the Clawd icon.
  • / menu: with Clawd on, the composer lists Claude Code's slash commands and skills (asked through the control protocol's initialize, no model call), filtered as you type; picked commands stay where they were typed and listed ones show in the primary color. A local command's output (/compact) becomes its reply and keeps the session.
  • One session, chat and window: a Clawd button under the browser button shows the agent window running the chat's session (resumed there when it is back at its shell, else a new window that records the session as its own), its border flashing once in the primary color. The chat and the window take turns in that session, so neither branches it: while the chat replies the window is read-only (with Take over) and resumes the session afterwards; while the window works the chat doesn't send, edit or regenerate, and takes the window's whole turns in from the transcript. "Open as chat" opens the chat the session belongs to.

Chats and agent windows message each other

  • Stable names: every chat (chat-…) and agent window (agent-…) has a name with a copy button. Both get Workbench's list_sessions and message_chat tools; windows are reached with Claude Code's own SendMessage.
  • Paste a name to address a message: in the chat composer it becomes an attachment (Claude icon for chats, Clawd for windows; the latest paste wins); in an agent window it is typed as [Chat #…] / [Agent #…].
  • Closed windows still count: messaging a closed (or not yet started) window opens it again in the background on its last session, in the folder it ran in (a worktree too), so it answers even while you are on another page. A window back at its shell resumes its session, but only when nothing else runs in it. Windows accept messages from your other sessions unless your Claude Code settings refuse or hold them.
  • Read-only while answering: a window a chat messages ignores typing (not the terminal's own replies) until it has answered, with "Answering … · Take over" in its header.
  • Open as chat: an agent window's header opens its session as a chat (disabled while Claude works); opened again, the same chat shows what was said since.
  • Safety: a chat answers other sessions on its own at most 5 times in a row, and those replies run read-only (no Clawd, Kanban, GitHub or GitLab tools). Messages waiting for you are shown without a reply and passed on with your next message. A read-only chat gets SendMessage only when the turn reaches an agent window.
  • Messaging tools have their own icon in the chat's tool list.

Claude page

  • Opens as you left it: the last chat of each project (also after a restart), the composer's text and recipient, the history panel and the chat's browser drawer view.
  • Sidebar's Claude item on the Claude page starts a new chat, like the New chat button.
  • Width follows the composer: the chat's narrowest width is what its composer's tool row needs (measured when a chat opens or its composer, model label or font changes), instead of a fixed 500px. Widening the browser drawer closes the chat history, then collapses the app's sidebar, and stops there; they open again the other way round. Dragging the drawer renders only its frame.
  • Messages wrap to the chat's width (a long tool command made them overflow); the new chat's send button shows disabled until there is something to send.

Browser drawer and chat browsers

  • Settings > Browser (new page): the default browser profile (moved from Connectors), the homepage a new chat browser opens on, the bookmarks bar switch and the extensions switch.
  • Like a real browser: the profile's bookmarks bar; links open in new tabs with Ctrl/Cmd+click or a middle click; tabs the user opens are shown and focused, Claude's stay in the background; the browser's own shortcuts (Ctrl/Cmd+T, W, Tab, 1-9, L, R, Alt+Left…, letters as the layout types them) act on the drawer's tabs; copy and cut reach the system clipboard; pages show their own cursors; Back and Forward follow the tab's history and are disabled with nowhere to go; reload turns into Stop while a page loads.
  • Extensions: chat browsers keep the profile's extensions, shown at the address bar's end (pinned ones, then a menu with a switch per extension; the menu is hidden when empty). Clicking one opens its popup over the page, sized to its content and laid out at that width, closed by a click beside it. Chromium runs the toolbar action itself; Firefox opens the popup's page behind the tab, and where newer Firefox refuses to capture or drive extension pages over WebDriver BiDi, the popup captures itself (tabs.captureTab) or is drawn from its own document, with input replayed as DOM events. Firefox's back-forward cache is off so Back and Forward report the page shown.
  • Loaders: a clicked extension's popup shows at once with Workbench's loader until its content draws; a loading tab shows the loader over its page.
  • Tabs kept per chat: closing a chat's browser (the drawer's close button, now asking first in an overlay on the drawer; the app quitting; the browser going away) keeps its tabs, and they open again with the chat's next browser. The browser button stays in the chat's top bar and toggles the drawer.
  • Homepage only when nothing else opens: a browser opened for a turn's browser_navigate opens that address in its first tab, without a homepage tab beside it.
  • The chat's Playwright can't open the browser's own pages (about: but about:blank, chrome:, resource:, extension pages).
  • Fixes: a viewed page that stayed blank starts streaming again; a tab switched to shows at once in Firefox; a Chromium page shown again after another tab was in front no longer shows scaled; Chromium profile copies keep extensions working (no app-bound preference hashes, service worker script cache kept); profiles only a chat picked keep their saved copies.

Windows and WSL

  • On Windows, chat browsers always run natively, also with the WSL shell (runNatively() runs code as a native session inside a WSL one). When Node.js or Playwright's browser build is missing on Windows, the drawer says what and offers Autoconfigure (a PowerShell window installs Node.js with winget and fetches them), then Open browser comes back without a restart.
  • With WSL mirrored networking, the chat's task tools and browser relay go through the stdio bridge (the Hyper-V firewall blocks Windows' loopback), and the task tools turn off when the bridge fails.

Accounts and usage

  • Automatic account switching removed: runs stay on the chosen account; a rate-limited task waits for its reset and resumes on it (only a removed or signed-out account is still replaced).
  • Delete Claude accounts from their row (on hover): signed out with claude auth logout, config dir removed (for ~/.claude only its sign-in settings), project references cleared. The app's account in use can be deleted from a window on another account, which takes its place.
  • One "In use" label: while an API profile is in use, no Claude account is marked.
  • API profiles: the key is optional for http://localhost (any port); the default model is required and Haiku/Sonnet/Opus follow it until changed. The active profile has no delete button.
  • Enterprise accounts: their monthly spend limit shows in the badge, the usage popover and the Accounts row instead of empty 5-hour and weekly limits (with a Simulation preview).
  • Stored usage keeps its reset times and the weekly Fable limit; usage fetch failures are written to the app's log.
  • An app started from a Workbench terminal ignores the OAuth token that terminal injects (it made every account show as "OAuth token").

Projects and windows

  • Pages as left: each project's window opens on the page it showed last, also after a restart.
  • Projects page: the latest opened projects first (the sort button goes on to A–Z and Z–A).
  • Closing a window: the only window hides to the tray as before; with other project windows open, a small dialog asks whether to keep it running in the background or close its project.
  • Agent windows: Claude Code redraws in windows shown again after the Projects page (they stayed blank); no expand button while only one window is open; working vs idle is read from Claude Code's current footer.

Smaller changes

  • Kanban: columns fit the board without a scrollbar (the Agents page's spacing, a lower minimum column width).
  • Live agent graph: a session shows in the window Workbench started or resumed it in, a chat's session too.

Tests

  • The browser-profile lock tests fake browsers with bash through a symlink: on Ubuntu 26.04 (the CI runner) /bin/sleep is Rust coreutils, which dispatches by argv[0], and Node renames its main thread.
  • The conflict check's temp folder removal is retried.

🤖 Generated with Claude Code

ArtyomVancyan and others added 21 commits November 7, 2026 22:12
Runs stay on the account the user chose: no switch on rate limits, usage
thresholds, capacity or auth failures. A rate-limited task waits for its
account's reset and resumes on it. Only an account that is removed or
signed out is still replaced, since nothing can run on it.

Removes the settings rows, the auto-switch settings and IPC, proactive
swaps in the usage monitor, reactive swaps for tasks and terminals, the
profile scorer, the swap notification and the unused queue API.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- A base URL starting with http://localhost (any port) can be saved
  without an API key; such a profile sends a placeholder credential to
  Claude Code and the connection test, which need one.
- The default model is required. Haiku, Sonnet and Opus follow it while
  empty or still set to it, until another model is picked for them.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- An account's Details dialog can delete it after a confirmation: it is
  signed out with `claude auth logout` and its config dir removed (for
  the default ~/.claude only the settings that set up how it reaches
  Claude); projects that chose it lose the reference. The account in
  use, the only account, one with running tasks and a default-dir setup
  defined by the environment or managed settings can't be deleted.
- While an API profile is in use, no Claude account is marked in use;
  using a Claude account switches the API profile off.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The main process build doesn't resolve the @shared alias for runtime
imports, so `npm run dev` failed to start.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…inal

An app started from an agent window inherited that window's
CLAUDE_CODE_OAUTH_TOKEN, which account detection read as a machine-wide
setup overriding every sign-in: accounts showed "OAuth token" without
usage and lost "Sign in again".

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude accounts get a trash button before their menu, shown on hover
like API profiles' buttons, instead of a Delete button in the Details
dialog. The account or API profile in use shows no delete button
instead of a disabled one.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Use the Agents page's spacing (12px from the edges, 8px between
columns), lower the columns' minimum width to 180px so all six fit
without a scrollbar, and inset cards from the column sides as much as
from its header.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ive path

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
With the button on, a chat's turns run without the read-only tool list
or --strict-mcp-config (every tool, the user's own MCP servers too) and
with the agent windows' permissions: skipped when Settings skips them,
else auto mode, since a headless turn can't ask. Turning it off keeps
the same Claude Code session; later turns are read-only again. The
system prompt says which mode the turn is in, and tools only a Clawd
turn can call show the Clawd icon.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A Clawd button under the browser button resumes the chat's session in a
new agent window of the project (`claude --resume` on the account it
ran on), like a task's "Open agent". Shown once the chat has a session;
disabled while it replies or when the project has no room for another
window.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
With the Clawd button on, a `/` at the start of the input or of a word
opens a menu of Claude Code's slash commands and skills in the project,
filtered as the user types (arrows, Enter or Tab to pick, Escape to
close). They are asked once per project with the control protocol's
initialize request, which needs no model call. Also documents the Clawd
button in CLAUDE.md.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- A local command's output (/compact) is its reply and keeps the
  session, instead of failing as an empty response; the menu leaves out
  built-ins that don't fit a chat (/clear, /model, internal ones).
- A slash command folded into a new session's prompt stays at its start,
  where Claude Code runs it.
- The commands probe runs through startClaudeProcess (no uncaught EPIPE,
  stderr read after close), without MCP servers, and fails with codes
  the renderer words.
- The list is kept per project and account, asked again when the Clawd
  button is switched on and after a failure.
- Enter sends a command typed out in full, Shift+Tab and IME keys are
  left alone, the pointer only picks rows when it moves, and the menu
  is a portal placed above or below the composer so nothing clips it.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
With the Clawd button on, the commands and skills the `/` menu lists
show in the primary color where the message names them (a path or an
unlisted word starting with `/` stays plain). A copy of the text behind
the field draws them; the field's own text is transparent while it
holds one.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
On a busy CI runner the helper's 1s wait for a spawned process to run
under a browser's name ran out, and the test failed later in an
unrelated assertion. Wait up to 10s and fail in the helper instead.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Ubuntu 26.04 (the CI runner) ships Rust coreutils, one binary that picks
its tool by argv[0]: started through the `chrome` link it exited at once,
so the lock's pid was dead when checked. Start it with argv[0] `sleep`;
the process is still named after the link (comm), which the check reads.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
On Ubuntu 26.04 (the CI runner) /bin/sleep is Rust coreutils, one binary
that picks its tool by the name it runs as: started as a browser's name
it never ran as one, also with argv[0] `sleep`. Node started through the
link waits under the link's name on any Linux.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
While a page loads, the reload button becomes an X that stops it (CDP
Page.stopLoading, else window.stop()); the bridge tracks each page's current
navigation so a replaced one doesn't end the loading state.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Rust coreutils dispatches by argv[0] and Node renames its main thread, so
neither works as a renamed fake browser; bash run through a symlink does.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude Code's prompt is ❯ and its logo is made of blocks, so the old patterns
left windows marked busy. "esc to interrupt" in its footer means working; the
permission mode's "(shift+tab to cycle)" or a finished turn's "· done" line
means waiting for input.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The graph guessed a session's window by which one was typed into last, and a
window opened in the background was never typed into. The window that started
or resumed a session now decides first.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- Stable names (chat-…, agent-…) with a copy button on each chat and window;
  list_sessions and message_chat tools for chats and windows; a chat delivers
  messages to itself and auto-replies to at most 5 in a row
- Pasting a name sets the recipient: an attachment in the chat composer,
  [Chat #…] / [Agent #…] in an agent window
- A message to a closed (or not yet started) window opens it again in the
  background on its last session; the window is read-only while it answers,
  with a Take over button
- Windows accept messages from the user's other sessions through their own
  --settings file (which also names the theme)
- "Open as chat" on agent windows, disabled while Claude works; messaging tools
  get their own icon; list_sessions picks the process under the window's shell

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@ArtyomVancyan ArtyomVancyan changed the title Clawd mode for Claude chat, account deletion, no auto-switching Clawd mode and session messaging for Claude chat, account deletion Nov 8, 2026
ArtyomVancyan and others added 8 commits November 8, 2026 22:36
…ch other

- A read-only chat turn gets SendMessage only when it reaches an agent window
  (the user's message goes to one, or answers one) or with Kanban or Clawd on;
  unattended replies to other sessions run read-only, even with Clawd on
- Opening the target window runs inside the turn: Stop, newer messages and
  incoming ones see the chat replying, and Stop ends the wait at once
- Messages that arrived during a stopped or failed reply are shown, not lost;
  held messages resume the session and reach Claude with the next turn
- Windows reopen in the folder their session ran in (worktrees), an open window
  back at its shell resumes its session, and a window not started yet starts
  in the background at a usual size
- Read-only ignores only the user's keys (terminal replies and the app's own
  writes go through), and only for windows that accept messages without asking
- Open as chat keeps an archived chat unarchived and leaves out /compact
  summaries and interrupted-request notes
- Session lookups use the window's own account, share a running
  `claude agents` listing and cache reads; SendMessage and ListAgents labels

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- The page opens on the chat it showed last (per project, kept across
  restarts), or the New chat page when that is where the user left; a new chat
  when the chat is gone
- The composer's text and pasted recipient, the history's open state and the
  chat's browser drawer view are kept while the user is on other pages
- The sidebar's Claude item on the Claude page starts a new chat, like the
  page's New chat button
- Page storage is flushed on quit, so a quick quit keeps the latest of it

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ging

- Unattended replies to other sessions run without the Clawd, Kanban, GitHub
  and GitLab tools; a message arriving while others wait for the user waits too
- Held messages from other sessions are passed on once with the next turn (not
  also in its history), and a turn passing on an agent window's message may
  answer it; a turn after the user's own failed one starts a new session
- Messages waiting for a chat archived during a stopped reply are kept in it
- An open window back at its shell is resumed only with the shell in the
  foreground; a window whose shell exited opens again as a closed one
- A window without a recorded session is also looked up under the account its
  Claude Code was started on
- Read-only passes mode, keyboard and window reports, and a cursor report only
  right after the program asked for one (Shift+F3 looks the same)
- One helper shows incoming messages; tests restore stubs and timers in afterEach

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- Each project's last chat is read afresh before writing, so the app's windows
  don't overwrite each other's entries
- The history closed by the browser drawer isn't remembered as the user's choice
- The page lists the history and opens the last chat at once, and lists it once
  when that chat is gone

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
On Windows a git process the test started can still hold a file in it for a
moment, and the removal failed with EPERM.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The drawer's browsers run headless, so their own toolbar never shows. The drawer now draws the
profile's bookmarks bar under the address bar: Chromium's Bookmarks file, Firefox's places.sqlite
(read from a copy, since a running Firefox locks it). Folders open as menus, bookmarks that don't
fit go behind a button at its end, and a bookmark opens in the tab shown.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Settings > Browser, between Project and Connectors: the default browser profile (moved from the
Connectors page's Playwright row; tasks and chats still pick their own), the homepage a new chat
browser opens on (best effort: a page that doesn't load leaves it blank) and a switch for the
drawer's bookmarks bar.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
ArtyomVancyan and others added 26 commits November 13, 2026 16:58
The relay leaves headless sessions (a Claude chat's `claude -p` turns) out
of every window's graph, so a chat's session opened in an agent window kept
its graph waiting for a session forever. A headless session counts once a
window's hooks report it runs there, and is looked for at once then.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- The chat's Clawd button shows the agent window running its session (resumed
  there when back at its shell) instead of opening a new one each time; a new
  window records the chat's session as its own, so it is found again and
  restored. "Open as chat" opens the chat the session belongs to
- The chat and the window take turns in the one session, so neither branches
  it: while the chat replies, the window is read-only (with Take over) and
  resumes the session afterwards with the chat's turn; while the window works
  (`claude agents` status, which the output alone misses), the chat doesn't
  send, edit or regenerate, and takes the window's whole turns in from the
  transcript when it goes on

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Ctrl+W on AZERTY is Ctrl+Z (undo), so letters come from the key the layout types, falling
back to the key's place for non-Latin scripts as Chrome does. A held Ctrl+W, T or R acts
once (its repeats are swallowed), while switching tabs keeps going; a page that reloads or
crashes stops the catching until its drawer asks again.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- The window is checked as the reply's first step (the executor's `prepare`), so the chat
  already counts as replying: a busy window ends the turn with agent_window_busy, the
  user's message goes back to the composer and another session's message waits
- Only windows that recorded the session are looked at; one counts while `claude agents`
  says it runs that session, busy unless it reports idle
- Each reply holds the window read-only on its own; it resumes the session once no reply
  holds it, and a Claude Code exit in the window keeps the holds
- The window's turns come in one sync at a time per chat, whole turns only, before the
  messages the chat holds, re-reading the transcript only when it changed
- The Clawd button resumes at a shell only with nothing else running there and ignores
  a double click; the chat checks the window again before a send

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The window a chat's Clawd button shows, open already or new, eases its border into the
primary color and back, so among several windows the chat's session stands out.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ects page

The Projects page unmounts the agent windows, and coming back resizes each PTY to the size
it has, which is meant to make Claude Code redraw: one column narrower, then back. Both
resizes reached Node before it handled SIGWINCH, so it read an unchanged size and never
redrew, leaving the window blank until another window changed its size. The narrower size
now holds for 200 ms before the real one comes back.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A window showing a project records when it was opened (Project.lastOpenedAt). The Projects
page lists the latest opened first and shows that time; the sort button goes on to A–Z and
Z–A.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…w closes

While the tray keeps Workbench in the background, the close button asks first: keep running
in the background hides the window to the tray as before; Quit quits Workbench, or closes
only this project's window when others are open. A renderer that can't ask hides the window.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…owser drawer

- The chat's narrowest width is what the composer's tool row needs (each part at its full
  width, the disclaimer truncating first, scrollbar room included), measured when a chat
  opens or its composer, model label or interface font changes, instead of a fixed 500px
- Widening the drawer, the chat needing more or the window narrowing closes the chat
  history, then collapses the app's sidebar, and the drawer stops there; they open again
  the other way round. useChatRoom works it out from the stores outside React, also when
  the page shows with the drawer open
- Dragging the drawer renders only its frame (BrowserDrawerFrame), not the chat page or the
  drawer's content, and saves its width once let go (or cancelled)
- The new chat's send button shows disabled until there is something to send
- The panels beside the chat take their widths from one place (lib/layout-widths.ts)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- A chat browser's tabs are kept when it closes (the drawer's close button, the app
  quitting, the browser going away) and open again with the chat's next browser; the
  drawer says how many. A deleted chat, a last tab closed, or a chat gone at startup keeps
  none. Kept in <userData>/chat-browser-tabs.json, read in the background, written whole
  and owner-only; kept until the browser has started, with tabs still reopening counted
- The drawer's close button asks first (an overlay on the drawer) and closes the chat's
  browser; with none open it only hides the drawer. The chat's browser button stays in its
  top bar and shows or hides the drawer
- A browser opened for a turn's browser_navigate leaves its first tab to that address, also
  when the call comes while it starts: no homepage tab beside it. Opened with nothing to
  open, it shows the homepage as before

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… are open

The only window hides to the tray when closed, as before. With other project windows open
the close button asks whether to keep this one running in the background or close its
project; closing it never quits Workbench, even when the others closed meanwhile.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…in the drawer

- The extensions button and the fallback extension icon use Workbench's own
  extensions icon instead of lucide's puzzle
- The drawer's page shows the cursor the page asks for where the mouse is
  (pointer over links, text over text...): after each pointer event the bridge
  asks the page (looking into same-origin and cross-origin frames) and sends a
  changed cursor to the windows viewing it, popups included

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…working

- A copy leaves out the encrypted hashes of its protected preferences
  (`*_encrypted_hash` in Secure Preferences). They are app-bound like cookies
  and never match in a copy, so Edge and Chrome reset those preferences and
  deleted the extensions they no longer knew. The plain hashes beside them
  hold on the same machine
- A copy keeps `Service Worker/ScriptCache`: without the scripts, extensions'
  service workers were dropped and never registered again, so their popups
  waited forever for a background that never answered

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…as the browser's own

Clicking an extension opened an empty popup, and its content didn't fit.

- Firefox over WebDriver BiDi (the user's own Firefox): BiDi refuses to open
  `moz-extension://` pages unless system access is allowed, so chat browsers
  running extensions start with MOZ_REMOTE_ALLOW_SYSTEM_ACCESS=1 (the registry's
  `moz-firefox` channel tells the bridge it is BiDi). BiDi still refuses to
  resize, activate or report events of such a page: the popup loads in front,
  hidden, to be measured, then loads again behind the page's tab, so the
  extension acts on that tab; closing it first leaves it for about:blank, else
  Playwright keeps it. Juggler keeps the fixed-size, mouse-only popup tab
- Chromium: a headless browser's popup never grows to its content and takes no
  other size, so the popup the browser opens gives its address and closes, and
  that page opens in a tab behind the page's, measured to its content
- Popups are sized to their content (its preferred width and its height at
  that width, at most 800x600) and keep their tab's own viewport, of which the
  drawer shows the corner the content takes, without a border or background
- A popup closes on a click anywhere beside it, the window losing the focus or
  its button again; another extension's replaces it

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… front

In headless Chromium, a page whose tab was hidden while its screencast ran (an
extension's popup tab, a tab Claude opened) and is shown again sends frames of
the browser window's size instead of its viewport, so the drawer showed it
scaled until its viewport changed. Such frames are now left out and the
viewport is set one pixel wider and back, which brings the right frames back;
after three tries in a row frames pass as they come, so the drawer never goes
blank.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…rawer

- On Windows chat browsers always run natively, also while the shell is WSL:
  the Browser runtime row (Settings > Browser) and its startup decision are
  gone, Windows' Node.js is looked for when a browser opens, and the code that
  ran them inside the distribution (the bridge there, the registry read through
  \\wsl.localhost, the bridge's endpoint probe) is removed. With WSL interop
  off, chat turns open a browser of their own in the distribution as before
- While opening the picked profile needs something missing on Windows (Node.js,
  Playwright's own browser build), the drawer's empty state says what and offers
  Autoconfigure instead of Open browser: a PowerShell window (Start-Process from
  a hidden PowerShell that waits for it; a detached start from the app got no
  console, so the script never ran) installs Node.js with winget (also found by
  App Installer's path) and the build
- After Autoconfigure, what is missing is checked every 5 seconds (up to 15
  minutes) and when the window closes, so Open browser comes back as soon as
  the browser can open, without a restart; the button reads Installing... meanwhile
- A failed open is logged, makes what is missing be checked again, and its
  error is kept across the drawer's loader instead of disappearing

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Expanding a window only means something beside other windows: with one open
(of any kind) its expand/collapse button is hidden, and a window left alone
isn't kept expanded for when another opens.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…failures are logged

After an account switch, or while a fetch fails, the popover shows the
account's stored usage. It had only the percents, so the 5-hour limit read
"starts when a message is sent" and the weekly Fable row was gone.

- Each fetch saves the 5-hour and weekly reset times and the model-scoped
  weekly limit (percent, reset, model name) with the account (the active one,
  those fetched in the background and those sharing a login), and the stored
  usage shows them
- Usage API errors, 429 back-offs, network failures and failed checks are
  written to the app's log (they went to the console, which packaged builds
  don't keep)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…Playwright's output folder

- Project memory: browser drawer features should act like the browser's own,
  through native mechanisms where there are any
- .playwright-mcp/ (Playwright's snapshots and screenshots) is ignored

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A chat browser running extensions is Firefox with system access over WebDriver
BiDi (so the drawer can open extensions' popups), where the browser's own pages
run with its privileges. The chat's Playwright guard now refuses to open them:
browser_navigate and browser_tabs (new) may open http(s), file, data and blob
addresses, about:blank and addresses without a scheme, but not about:config,
chrome:, resource:, extensions' pages and the like. Only Workbench's bridge
uses that access.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… stay light

- A popup shown from a tab is laid out at the width it was measured at
  (its <html> is given that width), as the browser's own popup is: it spread
  over the whole tab before, so a full-width header or centered content was cut
  off in the corner the drawer shows. Firefox's popup reloaded behind its page's
  tab gets the width again once it answers
- A Firefox popup whose page never answered is loaded again behind the page's
  tab all the same, so the extension acts on that tab and not on the popup's own
- The main process waits up to 45 seconds for a popup (a Firefox browser's first
  extension page loads twice), instead of reporting it failed at 15
- Clicking in a popup measures it again in one loop, a newer one taking over
- Frames of the old size right after the drawer resized a Chromium page pass as
  they come instead of setting its viewport again
- Cursor lookups look into open shadow roots, find an iframe's frame in one call,
  and skip popups whose pages never answer (Juggler's, and BiDi's once behind
  their tab), so no lookups stay pending

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- With the WSL shell, Playwright's availability check runs in the distribution
  while chat browsers run on Windows: it no longer disables the drawer's Open
  browser (the Windows check stands in for it, `fromWsl`)
- With the PowerShell shell, Node.js is looked for as runs find npx
  (resolveCommand), so a Node.js only fnm or the login shell knows no longer
  keeps offering Autoconfigure
- A chat's open error goes once its browser opens, also when a chat's turn
  opened it
- Leftovers of chat browsers run in WSL are gone: chatBrowserWsl(), the
  drawer sessions' WSL path mapping, the distribution's temp folder for
  holders (the startup sweep still looks there); the Playwright build a choice
  needs comes from one mapping

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…irefox

Firefox 157 refuses to capture or drive an extension's page over WebDriver BiDi ("browsing
contexts in privileged scope"), so popups stayed empty on Linux.

- A popup loads once, behind the page's tab (the extension acts on that tab), and is never
  reloaded: its script context stays, which the rest needs
- When the browser refuses to capture it, the popup captures itself with its extension's
  tabs.captureTab (the browser's own rendering; Firefox blurs other extensions' pages). One
  without <all_urls> is drawn from its own document: sheets with imports and media queries
  resolved, resources inlined, form state, scroll, canvases, sprites, same-origin frames, shadow
  trees with their computed styles, frozen animations, native toggles drawn
- Its input is replayed in the page as DOM events, into shadow trees and same-origin frames:
  clicks, the caret where clicked, typing and editing keys through the editing commands, Space,
  cut, the wheel
- Captures that keep failing switch to drawing; an unchanged popup is captured less often; one
  paint at a time, repainted only after a change or while animating

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…pty extensions menu

- A clicked extension's popup box shows at once with Workbench's loader, until the popup the
  browser answered with is listed; its own box keeps the loader until its first frame is drawn.
  A click beside it gives up waiting, and the page under it still takes clicks
- While the tab shown loads, the loader shows over its page
- The extensions button hides while it would list nothing (every extension pinned), unless a
  restart waits in its menu; its menu doesn't reopen on its own when something is listed again

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- Firefox's back-forward cache is off for profile copies: a page it restored came back without
  WebDriver BiDi's navigation events, so the drawer kept the address and title of the page left
- Back and Forward are disabled with nowhere to go: the bridge follows each tab's history as it
  navigates (its address against the ones before and after, history.length telling a new entry
  from a replaced one, a tab's starting blank page not counted), whoever navigates

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The page a project's window shows is kept per project in localStorage
(flushed on quit), and a window opens its project on that page, also after
a restart, instead of the Kanban board. A Memory page whose graph is gone
falls back to Kanban as before.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@ArtyomVancyan ArtyomVancyan changed the title Clawd mode and session messaging for Claude chat, account deletion Clawd mode and session messaging for Claude chat, a fuller browser drawer, accounts and window improvements Nov 14, 2026
@ArtyomVancyan
ArtyomVancyan merged commit 138c1e6 into main Nov 14, 2026
3 checks passed
@ArtyomVancyan
ArtyomVancyan deleted the develop branch November 14, 2026 17:24
@ArtyomVancyan
ArtyomVancyan restored the develop branch November 14, 2026 17:24
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant