Skip to content

Web UI Guide ​

The CCCC Web UI is a mobile-first control plane for managing your AI agents.

Accessing the Web UI ​

After starting CCCC:

bash
cccc

Open http://127.0.0.1:8848/ in your browser.

cccc is the single owner of the default local app session: it starts the daemon and Web together, and pressing Ctrl+C stops both together. If another cccc session is already running for the same CCCC_HOME, a second cccc command will refuse to start instead of silently sharing the old daemon.

If the Web server and its daemon report different versions, a dismissible notice shows both and says that some features may not work. A difference does not by itself establish that they are incompatible. If version information is unavailable, the notice is not shown. cccc daemon start, cccc web and cccc --headless also report a different running daemon version when they reuse one; they do not replace that daemon. Dismissing the notice remembers that pair of versions in this browser's site storage; a different pair can show a new notice. If storage is unavailable, dismissal lasts until the page reloads.

If Waiting for you says that the running daemon does not support it, finish current work before restarting: on the CCCC host, use the same CCCC_HOME to run cccc daemon stop, then start CCCC from the intended installation. Restarting only the Web can leave the old daemon running. If services are managed separately, use their service manager to restart them as appropriate; remote users should ask the administrator. Stopping the daemon can interrupt work still in progress.

Interface Overview ​

The Web UI has these main areas:

  • Header: Group title and controls, work view switch, and settings menu
  • Sidebar: Group list, navigation, and the persistent Voice control
  • Main Area: Group message history or a paginated terminal view
  • Input: Message composer with @mention support

The local Group list scrolls independently from the collapsible CCCC Connect section and the Voice dock below it. CCCC Connect starts collapsed; click its heading to expand or collapse the instance list. The section's information button opens the full local instance name and public address. Group rows are more compact, while their ordering stays unchanged; connection counts and action menus keep fixed positions, with menus visible on touchscreens.

The composer uses one send control with an adjacent mode arrow. The primary action shows the selected mode; its tooltip includes the send shortcut. Recipient chips need no separate “To” caption. A message within the Group names its recipients in the header, and one addressed to you shows You (in a read-only exhibit, User); a message sent to another Group or instance shows its destination route. User names remain accessible to screen readers without repeating a visible label above right-aligned local user messages. Message spacing is tighter, and the composer uses a subdued border that strengthens on focus.

Image attachments in the composer show thumbnails before sending. Hovering or focusing a thumbnail opens a larger preview; the filename remains its accessible label and tooltip. Unsupported or unreadable images fall back to a filename chip, as do ordinary files. The attachment area scrolls when needed and local preview URLs are released when a file is replaced or removed.

While a message with attachments is uploading, a stop button appears next to Send. It ends the wait and keeps your draft: it returns to an empty composer, or waits to be restored if you have started a new one. It does not recall the message, which may already have arrived, so check the conversation before sending it again.

On wider screens, an empty, unfocused composer cycles through four localized typing hints for tasks, Skills (/ and /install), agents (@), and Groups (#). Focusing the composer pauses the hint; any draft text hides it, and a suggested reply takes precedence. Reduced-motion preferences show a static hint, and hidden tabs pause the animation. Motion set to Off under Appearance also shows a static hint. Hints are decorative and never become draft content or screen-reader live announcements.

Each message is as wide as its content, header and attachments included, up to the column; a question waiting for your answer keeps room to answer in. Each run of messages from one sender opens with a thin line that spans its first message: the first message loaded, a message after a change of sender, and a message after a notice or any other entry that is not a message. Consecutive messages from the same sender, an agent's or yours, are set apart by space alone. In a wide reading column, agents' messages and yours keep the same room from the opposite edge; in a narrow one, as on a phone, beside the files panel or at a large text size, agents' messages may use the full width. Space inside a message is reduced so that it reads apart from the space between messages, and a Markdown rule inside a message is a short mark in the text column. Images and files sent with a message follow its text with no rule between them, and the sender's insight follows both. Messages have no background, and pointing at one changes only its actions. Sender, recipients, Mail mode and time share a compact header; delivery failures and reply/read status stay visible. On a phone, the sender, time and actions sit above the route, the message uses the available width, and the sender's insight is set apart below it. A reply shows up to two lines of the message it answers; clicking or tapping that quote, or pressing Enter on it, opens the original message. With a mouse or trackpad and enough room in the message column, Reply, Forward and a copy icon appear beside the message header on hover or keyboard focus, after the time or before your name, in the free space next to it, so nothing shifts; a long header wraps to leave them room. The icon copies the message content as the menu item did before. Quick reply, Pin message, Copy link and other actions are in the bar's More menu. On touch screens or in a narrow message column, More ends each message's header and lists every action, including copying. Read-only views retain status information without exposing these actions. The composer groups recipients, text and tools into one surface.

On phones, the chat keeps as much of the screen as possible for messages. An empty composer that is not in use is a single row with the attachment button, the message box, voice and Send, plus a short label for a chosen recipient; tapping it or one of its controls expands it. Scrolling the messages by touch toward newer ones slides the header, the filter row and the runtime dock out of view. They return when you scroll back, tap the messages, reach the newest message or touch the composer. A deliberate touch scroll of the messages can also take the focus out of the message box, so the keyboard can close; a draft stays, and text an input method is still composing is left alone. The More button sits beside each message. When both are visible, Back to latest sits beside the members at the bottom of the log. Phones in landscape also use the single-row composer.

The header uses its existing Group status dot for page-connection health: red means disconnected, pulsing amber means reconnecting, and a connected page uses the normal Group lifecycle color. The badge text still describes the Group's runtime state (for example, Idle); hover or focus exposes connection details. No separate connection subtitle or extra dot is shown on desktop or mobile.

Opening a Group or message link selects that Group. Otherwise, the browser restores its last selected Group. If that Group is no longer available, the UI opens the first unarchived Group in sidebar order, or the first Group in that order when all are archived. Switching Groups updates the current tab's URL and clears the previous message link, so refreshing stays in the Group you selected. Reordering or archiving Groups does not move an existing selection.

Group message and terminal views ​

Use Connect sessions above the Group list to admit user-owned Codex, Claude Code, Pi and OpenCode V2 CLI sessions. Their native applications retain runtime ownership; the Web shows Group access, receive availability and recovery guidance instead of a mirrored terminal. See native session setup and lifecycle.

On desktop, Group editing, search, context and runtime controls stay directly in the header. Group connections is available under the current Group’s settings and in each Group’s sidebar ⋮ menu, including remote Groups in CCCC Connect. The settings entry manages the current Group; the sidebar menu manages the Group whose menu you opened. Only incoming invitations ask which local Group should accept the connection. Settings and more opens with your account, where your access includes it, and the full settings entry, then Appearance: theme and language stay in view, and More appearance settings unfolds body font, message area width, motion, ‘Waiting for you’ indicators and text size. The menu opens folded each time. A choice applies with one click or tap and keeps the menu open; arrow keys move a choice's selection. Narrow headers use the overflow sheet with the same choices and access rules; message area width is offered only on screens 768px and wider. Message identities, recipients, references, timestamps and actions follow the text-size preference. ‘Waiting for you’ indicators under Appearance turns off what marks requests waiting for your answer: the sidebar's counts, including the archived Groups' total and the one on the phone's sidebar button, and the Waiting for you filter inside each Group. Those requests then read as ordinary messages: one without a question is answered with the message's Reply, and one that asks a question still shows it so you can answer in place, without the Waiting for you frame. An answer already being sent keeps its form until it is confirmed, so it can be retried. A question's options can take a note: after you choose one, + Add a note (optional) beside Send answer opens a box, and the requester receives your choice and your words in one reply. A blank note sends the choice alone; Other keeps its own box, and switching between them keeps what you wrote in each. Browser notifications keep their own switch under Settings → This instance → Notifications. The indicators show by default; the choice is saved in this browser. Message area width under Appearance sets how wide the conversation runs on screens 768px and wider: Medium keeps a centered reading column, Wide gives tables and code more room and Full uses the whole chat area. With Medium or Wide, the conversation's scrollbar sits beside the column, and the mouse wheel scrolls the conversation from the blank space on either side as well. The choice is saved in this browser. It does not appear on narrower screens, where the conversation already has no width limit. The members' status floats in a compact centered row above the composer. Text scrolled behind the member keys fades out, while messages remain visible and interactive on either side; the scrollable tail lets the final message move fully above the member controls. Activity and feedback move in whole steps: a busy member's ring of pixel dots, three blinks when the page sees a member start needing attention, a pixel spinner for waits longer than a quarter of a second, and a block cursor that can follow a headless agent's live text, though not after code blocks, tables and some other structures. Motion under Appearance stills these and the typing hint; the system's reduced-motion setting does the same, and Appearance says so. The Voice Secretary level meter keeps showing the live level. Reading controls share the application theme; embedded PDFs, web pages, images and terminal ANSI content retain their own colors. Image and diagram viewers zoom with their buttons, Ctrl/⌘ + wheel or a pinch, while the plain wheel keeps scrolling; see Inspect images and diagrams.

In Group settings, refreshing data or saving a different section preserves edited fields and updates fields you have not changed. Delivery, Messaging and Transcript display save results in the current settings view; a failed save leaves your changes available for retry. Notebook refresh also preserves a pending notebook choice. Changing Groups starts from that Group's saved settings. Closing settings still discards unsaved form edits. Web Access refresh preserves the selected access goal together with edited connection fields. Saving or resetting one Guidance document updates that document without reloading the other editor; Discard changes remains explicit.

The Group status button combines the status dot with run controls, without a dropdown arrow. When space allows, the header also shows the state text; narrow headers keep a compact button. The sidebar keeps a passive status dot and places run actions in each Group’s ⋮ menu and context menu. These actions target that exact Group without switching views. Start Group starts enabled Agents. Pause message delivery pauses CCCC delivery; work already in progress may continue. Resume running resumes delivery, starting enabled Agents if none are running. Stop Group stops running Agents while preserving their enabled settings. Pending operations disable run and delete menu actions to prevent duplicate submissions. Read-only views retain passive status indicators. On mobile, use the same header control; collapsed sidebar icons remain navigation.

Use Messages / Terminals in the Group header to switch the current Group's work view. On narrow headers, the view icon switches between these views. Presentation also opens from the header through the bookmark icon, beside the Files folder icon on desktop, with its own update indicator. Tooltips and accessible names identify these controls. View, search and context controls align with the main work area as the side panel resizes. Files and Presentation share one side column: selecting the other swaps panels; selecting the active one closes it. Switching views only replaces the history area; the message composer stays in place. Each Group remembers its view and terminal page in this browser, including after a page reload. Message drafts and the message reading position survive switching between views.

The terminal view shows up to four agents per page, in the Group's existing order. Narrow areas show one agent per page. Page arrows appear only when there is more than one page, in the Group header or the single agent title bar on mobile. The arrows have 44-pixel targets, with a visible page count and disabled endpoints. On touchscreens, swipe horizontally on the agent title area to page; buttons, vertical gestures and the terminal body keep their own behavior. Resizing keeps the focused agent visible, or preserves the first visible agent when no terminal has focus. Removing agents displays a valid page without discarding the saved preference during loading. Indicators can point to waiting or stuck agents on other pages without moving the current page.

Each visible terminal accepts input directly. Click its content to choose the keyboard target; the focused pane is outlined, and new output does not take focus. The bottom composer still sends Group messages, independently of terminal input. Maximize opens the existing agent controls using the same terminal instance; close the expanded view to return to the grid. Escape and Tab remain terminal keys while the terminal has focus. Existing access and single-writer rules apply: a read-only user cannot type. Opening a view leaves any existing writer in control; use Take control to take over explicitly. An ordinary reconnect does not take control away from another writer. With no existing writer, the terminal accepts input immediately. A reconnect that resumes contiguous output keeps the existing screen visible while catching up.

Switching pages, Groups, or back to Messages keeps recently visited terminals and their connections in memory, including scrollback and selection. Each workbench retains up to 32 hidden terminals for five minutes after leaving them; visible terminals do not count toward that limit. The oldest hidden views are released first, and background output does not extend their lifetime. Unvisited agents are not loaded in advance. Hidden terminals continue receiving output but cannot accept input, resize the remote terminal, or take control automatically. On return, a retained writer synchronizes the PTY with its visible dimensions, including when control returned while hidden. Another window can still use Take control. Expiry, a browser reload, or a replaced agent may require a fresh terminal view; retention does not start or stop agents.

Each agent title bar keeps its status dot, common terminal actions and maximize control together. Ordinary running/working states are available through the dot's label; stopped, waiting, stuck and terminal connection notices remain visible as short text, including in narrow panes. In narrow panes with pagination, the agent name and notice occupy their own row above the page and terminal controls, so touch targets do not crowd out the agent's identity. Use More for history, session, configuration, inbox and lifecycle actions. Unavailable actions are disabled; stopping and restarting use the same operations as the expanded view. Different agents have independent pending-operation indicators.

Once a retained view expires or is evicted, returning attaches to the agent's current output; this never stops or restarts agents. Agents without a TUI use their existing runtime activity display, and stopped agents are explicitly identified. Hidden Group messages do not clear the message unread count or count as viewed for Voice suppression. Following a message source link returns to Messages and locates the original event.

The Runtime Dock uses a thin violet-to-rose trailing arc for working agents, rotating gently once every 3.6 seconds. Idle agents have a quieter green ring, stopped agents gray, and states needing attention rose. All rings keep the same size; reduced-motion preferences leave the working arc static. Hover or focus an agent for its name and use the button's description for status details. No duration or progress is inferred from status timestamps. The upper-right envelope badge shows unread Mail from the daemon, including messages outside the loaded chat history. Browser delivery queues have a separate badge at the lower right; both badges appear only when their count is nonzero.

The Runtime Dock's progress bubbles show the latest short live update for each agent, with up to two agents visible at once. Updates from the same agent replace the previous excerpt. Open the agent to inspect the complete output. Restoring a Group or reconnecting its activity stream restores history without replaying it as new bubbles; ordinary message unread indicators are independent of these brief progress previews. The same rules apply to every managed Runtime.

Files and Presentation share one resizable sidebar with matching header controls. Each Group remembers its expanded width and Presentation density. Files always opens at the expanded width; its file draft survives switching to Presentation and back.

Presentation keeps four fixed slots. New Groups open a compact button rail by default. Open a filled slot to read it in the resource panel; empty slots open the pin flow. The expanded overview uses image thumbnails and compact titled entries for other resources, without starting interactive browser/PDF sessions or clearing update indicators.

Collapse returns to the compact rail; closing the sidebar only hides it and keeps its contents. Dragging to the minimum width also returns to compact slots. The divider supports arrow keys, and Enter toggles density.

Resource panel ​

Workspace files (from Files, a path in a message, or a link inside a document) and Presentation slots (from the slot list or a message's reference) open as tabs in one panel. Message attachments open as tabs when the desktop panel is visible. Opening something that is already open selects its tab. Each Group keeps its own tabs. Refreshing the same browser tab restores their order, selection, side-by-side layout, and whether the panel was open, collapsed, or maximized, when browser session storage is available. This also keeps the layout when you switch Groups and return. Renaming or moving a file or folder in Files takes its open tabs and unsaved edits along.

Restoration saves resource references, not document contents or unsaved edits. Resources are read again with your current permissions. A restored message reference that is no longer available stays unavailable; it does not switch to the slot's current card. A request that fails offers a retry instead. Attachments not yet confirmed by the ledger are omitted. Reading positions and recently closed tabs are not saved across refreshes. This feature does not copy state between browser tabs or guarantee recovery after closing a tab.

The panel sits in one of two places. Docked right, the default, it takes the right-hand column. Docked left, it covers the message list and keeps the composer. The arrow button in the panel's bar moves it, and this browser remembers the choice. Docked left, the panel belongs to the message view: the terminal view hides it and the message view brings it back. For an attachment tab shown alone, View full screen opens its preview over the page. With two resources shown side by side, or when viewing a workspace file or Presentation slot, Maximize covers the messages and the right column together. Press it again or Escape to restore. While the panel covers the messages, docked left or maximized, newly arrived messages stay unread, and the collapse key shows how many arrived. The resource button in the Group header shows how many tabs are open and opens or collapses the panel. An empty panel opens alongside the messages, where you can click an attachment to open it there. Escape collapses the panel only from its bar. Editors and PDF viewers keep their own Escape.

When the tabs do not fit, the tab strip scrolls sideways, also with the mouse wheel. All lists every tab in order with a filter by name, path or sender, followed by Recently closed. Several tabs closed at once come back together, from that list or from the Undo notice that follows for eight seconds. Closing the last tab leaves the desktop panel open with Recently closed available inside it. If Files or Presentation is open in the right column, it takes that space again. Focus the tab strip and use the left and right arrow keys to select open tabs. Right-click a tab for Close, Close others, Close tabs to the right and Close all. All also offers Close others and Close all. Delete closes the selected tab.

A file with unsaved edits never closes silently. Closing it asks whether to save, close without saving or cancel; closing several stops and offers to close only the saved ones.

Switching between workspace text, Markdown or table tabs in a Group, each shown alone, returns to where you were reading, so two long documents can be read in turn; a file opened at a link's anchor starts there. A workspace PDF or video starts again from the beginning when you switch to another workspace file, and an HTML preview does not keep its scroll position. While a file is loading, or if it cannot be read, its tab says so rather than showing the previous file. Attachments you looked at most recently stay loaded, eight images and four other attachments; an older one loads again when you return to it, so a PDF's page, a video's position or a document's scroll may start over. Switching to another Group, or resizing the window between the phone and desktop layouts, loads the panel again: tabs stay, but reading positions and loaded attachments start over.

Side by side shows two tabs next to each other, for comparing two documents or two pictures; on a phone they are stacked. It starts with the tab you are reading and the one you read before it, in tab order, or with the same tab on both sides when it is the only one open, and coming back to it finds the same pair while both are open. Each side's header names its tab and lists every open tab to show there instead; picking the other side's tab shows one document in two places, each side keeping its own position. The current side, marked in its header, is the one you last clicked or tabbed into (inside a page shown in a frame, such as a PDF in the browser's own viewer, click the header instead); a tab picked from the tab strip, All, a message or a link opens on it, and one the other side already shows makes that side current instead. A side keeps its place while it shows the same tab; a workspace file shown there again starts from the top, or at the anchor last followed there. Drag the divider to give one side more room, or move it with the arrow keys; a double click puts it back in the middle. Each side shows its resource as it reads: workspace files rendered or as text, with any unsaved edits, pictures, videos, PDFs and Presentation slots in their own viewers. To edit a file, leave Side by side with its key or Escape, and the current side's tab is shown alone. When both sides hold the same Presentation slot, its viewer stays on the current side and the other side says so.

A Presentation slot has one viewer in the panel however many of its tabs are open, and switching to another tab and back does not reload it. A slot's own tab takes the title of the slot's current card; a tab opened from a message keeps the title it was referenced with. Closing a Presentation tab closes only your view; the shared slot stays as it is. Open in window in the viewer shows the slot in a larger window and collapses the panel. The panel and the window share a slot's interactive page: when the slot is published again with another address while that page is open, CCCC asks before replacing it, since someone may be working in it.

On a phone the panel is a full-screen surface. Back returns to where you opened the resource from: the messages, the Files tree or the Presentation slots. All is where you close several tabs at once. To leave room for the Group title, the phone header has no resource button: the menu lists Resources with the number open, or, when every tab is closed, offers to reopen what was closed last.

Git diffs from Changes still open over the messages rather than in the panel. Enlarged message diagrams, and previews opened from a link in a message, such as an image or video on another site or Markdown from raw.githubusercontent.com, still open over the page.

Message search and settings ​

The conversation's Sender menu lets you check one or more senders. Each agent appears once; the current Foreman has a role label. Checked senders are combined with OR, and the tabs above the messages apply alongside them. User matches messages the user sent; the User-related tab also includes messages addressed to the user. Selections take effect immediately and the popup stays open for further choices. No checked senders means unrestricted. Show all and the sender's clear button remove only the sender condition; All resets only the tab. Each Group remembers the selected sender identities in this browser, so changing the Foreman role does not retarget a selection. Older matching messages share one history cursor and remain available through pagination. A removed agent's selected identity remains available to uncheck. Filtered history follows ledger append order and receives the same live read, reply and cancellation updates as the full conversation. Reconnect catch-up also covers messages beyond the bounded raw cache. Live arrivals are confirmed through a batched cursor read before joining the filtered rows, so an older replay cannot appear as the newest message. Failed reads keep a retry action; changing the filter or Group discards the departed view's pending reads. When a filtered list has earlier pages, Load older results above the messages reads another page, even if only a few messages are visible or a page has no matches. A short All list offers the same action when older messages remain.

When Waiting for you has loaded no requests and a fresh count confirms that none are waiting, the view returns to All. Requests answered or withdrawn during the current visit remain listed until you leave it. A failed load or an unknown or stale count does not trigger the switch.

Pinned keeps the messages you need to come back to. Choose Pin message in a message's menu and a small pin follows its time. The Pinned tab shows how many the Group has and lists them in full, in conversation order, each with Show in chat. A message unpinned there stays listed, so you can pin it again, until you leave the tab or come back to the page, when the list is read again. Pins belong to the Group and are kept by the daemon: everyone who uses the Group in this CCCC's Web UI shares them, on any browser or device, and each page reads them again when the Group opens or the page becomes visible. A read-only exhibit neither shows nor changes them, and agents do not receive them in messages or tools. A Group keeps up to 100 pins. When a pinned message can no longer be read, the tab says so and offers to remove its pin.

Changing a filter keeps your place. If you were reading earlier messages, the message you were reading travels with you: the last one you clicked, tapped or reached with Tab while it remains on screen, or the one across the middle of the list once you move on. If the new view contains it, it stays at the same height, as far as the view can scroll. If not, the first other message that was on screen and that the new view contains stays at its height instead. If none was, the nearest loaded message from that part of the conversation takes its place, shown from its top and briefly outlined. Otherwise the view opens at its newest messages, and your place is kept for when you return. If you were following the latest messages, the new view follows them too, unless you clicked or tapped a message that it contains: that message then keeps its height. A list too short to scroll always counts as following, so clicking a message there is what carries it over. A view you only look at, without scrolling or interacting with its messages, leaves your place unchanged, so switching back returns exactly where you were. Waiting for you and Pinned keep a message from your screen only when they list it, and otherwise open at their newest message. Messages that are not loaded, such as older ones a filter has not reached yet, cannot be kept. A place is kept for up to 30 minutes while the page stays open, not across a reload.

Sending keeps the filters. If they hide the sent message, a confirmation offers View message, including cross-Group sends; the action returns to the originating Group even if you have switched Groups. In a filtered view, Show in chat opens a message with the messages around it in the whole conversation. It leads the bar beside the message header where the column has room for it, or follows the message's time on touch screens and in a narrower column. Back to filtered conversation restores the filters and returns to the place you were reading.

Opening a message from a notification, a link or search shows it in the live conversation when it is already loaded and no filter is on; new messages keep arriving as usual. Otherwise Viewing a message shows it with the messages around it. Either way it opens just below the header and is held there briefly as the view settles, unless you scroll first; on a phone the Group list stays closed. Scrolling down past the end, or Load newer messages, loads the messages that follow, and scrolling on from the newest returns to the live conversation when no filter is on. Earlier messages cannot be loaded in that view. Return to latest at the top leaves it at any time, and the floating Back to latest button, when shown, also leaves it. With a filter on, the top button reads Back to filtered conversation, and both return to the filtered conversation.

Message search shows the current Group name and sender display names. Enter or Search submits the keywords; changing a filter reruns the submitted query. Typing new keywords does not change existing highlights or pagination until you submit again. Initial guidance, loading, no matches and failed requests are separate states. Open returns to the message context; reply and copying the event ID remain available.

Settings keeps This group and This instance separate. The ⓘ beside the switch explains both scopes, including why This instance may be locked, and says so when no Group is open. Section forms share a plain work surface, while independent configurations and authorization states retain their own boundaries. Under Branding, save the product name explicitly; icon uploads and resets apply immediately. The product name changes the Web sidebar, sign-in screen and browser tab, not the instance name shown in Connect.

Related fields use headings and spacing rather than nested cards. Section actions stay beside their heading when space allows and wrap on narrow screens. Settings help text follows the text-size preference. Switches support keyboard focus and Space, with a larger click/touch target around the compact track. These visual conventions do not change which settings save immediately or require a Save action.

Workspace files ​

Open Files in the Group header to browse the active workspace. The file tree and Presentation share the right-hand column. A file you open goes into the resource panel as a tab.

Save with the toolbar button or Ctrl+S / Cmd+S. Unsaved edits remain available when you switch files or collapse the resource panel, including internal symlinks to the same file. Switching Groups or instances with unsaved file edits asks whether to stay or discard; refreshing or closing the page uses the browser's unsaved-changes warning. If the active workspace changes, drafts stay with their original directory in this page. A tab from another directory shows its unsaved text read-only so you can copy it; switch back to that directory to edit or save it. A save never writes the draft into the newly selected directory. If the file changed on disk since it was opened, saving reports a conflict. Reload file reads the latest contents from disk, including updated media at the same path. With unsaved edits, it asks before discarding them; canceling or a failed read keeps the draft. Edits made while the reload is waiting are also retained. Reload is unavailable during a save. This check does not lock out external editors or agents.

The tree shows workspace files by default, including untracked and Git-ignored files. Git's own root metadata directory remains hidden. Use File browser options → Hide Git-ignored files to reduce clutter; this preference is remembered in this browser. Folders load only when expanded. Refresh directory refreshes the tree, preserving expansion and file drafts; Reload file refreshes the open viewer.

Paste a workspace-relative or absolute path into Go to file or folder, then press Enter. Files open in the viewer; folders expand and receive focus in the tree without closing your current file or losing its draft. Use . or the workspace root path to return to the top of the tree. Missing paths and paths outside the workspace report different errors. Reveal current file expands its parent folders and moves focus to its row; explicit reveal also clears the Git ignore filter so the target can be shown. Collapse all folders only folds the tree, keeping the file open. If you move on while a location is loading, its result does not take focus away from the editor, message composer, or another file selection. Use arrow keys to move through the tree and expand/collapse folders, and Enter to open.

Each row's … button, right-click, or Shift+F10 opens the same actions: attach the path as message context, copy a relative or absolute path, download a file, or pin a supported file to Presentation when permitted. Attaching inserts a path into the composer; it does not upload the file.

On desktop, File browser options → New file / New folder / Upload files / Upload folder works at the workspace root. The same actions in a folder's menu target that folder. File and folder menus also provide Rename, Move to…, and Delete. Move destinations include the full workspace-relative name, and their parent folder must already exist. Name collisions preserve the existing item. Deleting a folder permanently deletes its contents; the confirmation names the path and warns about affected unsaved edits. There is no recycle bin or undo.

Drag files or folders from your computer onto the Files tree to upload them; the highlighted folder and drop label identify the destination. Drop on the panel background for the workspace root. Native folder drops preserve subdirectories and empty folders. The folder picker includes the browser-provided file tree; browsers cannot include empty folders through that picker. If directory drops are unsupported, use the upload menu instead. Each selection is limited to 1,000 files/folders and 100 MiB total. Uploads stop at the first conflict or failure. Stop upload keeps completed entries; it does not roll back the batch. Git metadata cannot be uploaded. Dropping here does not attach files to your message.

Drag one existing tree entry from the current Files panel onto a folder (or the panel background) to move it within the workspace. Move to… offers the same operation without dragging. Moves and renames carry unsaved drafts to the new paths; a pending save must finish first. Changing a file's extension also updates its preview type while keeping unsaved text. A destination with another unsaved draft is rejected so both drafts are retained. Moving files does not rewrite imports or Presentation references. Refresh or repin those references when needed.

Files uses UTF-8 paths. A directory containing names that cannot be represented exactly shows an error; manage those names in the terminal. Names are never silently replaced with another file's path.

Select Changes next to Files to inspect the active workspace's Git changes. The list separates working-tree modifications, staged changes, conflicts and untracked items. A new folder is listed file by file, as code editors list it; it stays a single item when it holds a separate repository. It also stays one item when listing its files would exceed the time or size limit, and a note then says the list is incomplete. In Files, the contents of a new folder are marked untracked too. This also applies when the workspace itself is a new folder of its repository. Select a tracked change to see its diff in the main area; use Unified or Side by side when the viewer is wide enough. Untracked files open in Files, untracked folders reveal their contents, and conflicts open the current file for inspection. The view uses saved content; unsaved drafts remain intact when switching views. Refresh directory refreshes the current Git view as well. Lists are read on entry and after local writes while visible, without periodic background scanning.

Changes is read-only: CCCC does not stage, discard, commit or push on behalf of the user. These remain terminal operations because agents may share the repository's index. Large lists/diffs show an explicit limit; Git failures are not shown as a clean workspace. Phones retain their existing read-only Files behavior and can view Git changes in the same full-screen surface.

Symbolic links have a link icon. If the target is missing, outside the workspace, or inaccessible, its row explains why and disables content actions; you can still copy its path, or rename/remove the link on desktop without changing its target. CCCC does not repair links created by other tools or download files outside the workspace. After fixing a link on disk, use Refresh directory.

An empty file panel shows No files to display after loading completes. If ignored files are hidden, use Show git-ignored files to reveal them. Otherwise, add files to the workspace and use Refresh directory. Loading and failed requests have separate states.

A directory that cannot be loaded shows its error below the row. Use Retry there to load it again; an error does not mean the directory is empty.

On phones, Files opens a read-only viewer. Binary files and files larger than 1 MiB are not editable here. Files stay within the active workspace, and exhibit mode does not expose this surface.

Images open in the shared graphics viewer, including files larger than the text limit. Videos and audio use native browser controls, without autoplay, and pause when their view is hidden. PDFs open in the browser's PDF viewer, with Open in a new tab and Download file available if inline viewing is disabled or unsupported. Native support depends on the browser and the file's encoding. On a device whose main pointer is touch, such as a phone or tablet, and in a browser that reports no built-in PDF viewer, PDFs open in CCCC's own reader instead: the pages scroll, and Fit, Zoom out and Zoom in work as in the image viewer. The reader shows pages only: it cannot select or search text, follow the PDF's own links, fill its forms or open a file that requires a password. For those, download the file, or open it in a new tab unless CCCC is embedded through Connect.

Markdown renders with tables and Mermaid diagrams. CSV and TSV files open as tables, including quoted cells and multiline values; the preview shows up to 200 rows and 50 columns. Text-sized SVG, Markdown, CSV/TSV and HTML offer View source and Preview. Switching views preserves unsaved edits, and preview uses the current draft. The editor and the text view hold at most 1 MiB. HTML above that editing limit, up to 20 MiB, is read in full for Preview, while View source shows its first 1 MiB; the file stays read-only. Larger HTML shows its first 1 MiB as source only. Downloads always contain the saved file.

HTML opens as a static, isolated document: scripts, forms, nested pages and linked stylesheets are disabled. Inline styles and workspace image/audio/video resources are supported; it is not a running website or application preview. Pages keep their own size and margins: wide slides stacked in a centered column start at the left and scroll sideways. Markdown and HTML resolve relative file links within the opened workspace and open them in Files. Section fragments such as report.md#details or report.html#details position the loaded preview at that section, including repeated same-file links. Positioning happens once per navigation; ordinary updates do not pull the reader back. A PDF page fragment such as paper.pdf#page=5 opens that page, in the browser's viewer or CCCC's reader. HTML audio/video keep native controls whether the URL is on the media element or a nested source. HTML does not load external resources. Workspace paths, including percent-encoded names, still undergo the same permission and scope checks. Root-relative document links refer to the workspace root. Office documents and other unsupported formats can be downloaded; CCCC does not upload them to an online document viewer.

Media and PDFs stream directly from the selected workspace and support byte ranges; CCCC does not transcode or upload files to cloud storage. Each request checks the Group permissions and opened workspace identity. Remote playback uses the instance's upload bandwidth. Public Cloudflare Tunnel access remains subject to Cloudflare's video/large-file distribution terms; preview support does not establish an exemption or add a Cloudflare Stream subscription.

Inspect images and diagrams ​

Workspace images, expanded message images, Mermaid diagrams, Presentation images and enlarged quoted snapshots share a static graphics viewer. Use Fit, 100%, + or − to change scale. Fit never enlarges a raster image past actual size; rendered diagrams and workspace SVG previews can grow to fill the view. Ctrl/⌘ + wheel and a trackpad or touch pinch zoom at the pointer; the plain wheel keeps native scrolling and dragging pans enlarged content. A click on the content switches between fit and actual size; on touch screens, double-tap. See Message attachments for opening images and videos from a message. Focus the viewport for +/−, 0 (fit), 1 (actual size) and native arrow-key scrolling. Closing a modal restores focus to its opener. Regular message scrolling, PDF controls and interactive browser surfaces retain their own behavior. The workspace text editor is unchanged. Refreshing the same workspace-linked Presentation image preserves its zoom and position, including across a failed refresh and recovery. Image and Markdown updates replace content only after loading succeeds (and images decode). During a transient failure, the last loaded version stays visible with an update-failure notice. The next successful refresh clears the notice. Slow requests finish before another automatic refresh starts. First-load errors and confirmed missing or inaccessible resources show an error instead of stale content. Switching to another slot or publication clears the previous resource and starts images with Fit again.

Workspace-linked PDF and HTML readers retain their current page, zoom and navigation state. Use Refresh to load edits from disk; publishing a new card also updates the reader. They do not reload on the automatic image/Markdown refresh timer. On touch devices and in browsers without a built-in PDF viewer, a PDF that this CCCC serves, such as a workspace file or an uploaded one, opens in CCCC's own reader, as for workspace files; a PDF from another site keeps the browser's viewer. The slot viewer has no download; for a PDF the reader cannot open, Copy URL/path copies its URL or workspace path.

Chet global assistant (Experimental) ​

The Chet control below the Group list opens the global assistant, which is not a Group roster item. It shows Chet's pixel head, or an alert mark when something needs a look. The head blinks now and then and moves while Chet has work being sent, queued or carried out; with Motion off under Appearance, or the system's reduced-motion setting, it holds still. You can write a message without starting a voice call. Chet uses its configured, signed-in Runtime; if a Group is selected when you send, it is included as context, not as a permission boundary. The adjacent microphone/square control starts or stops live audio. Minimizing Chet does not stop its work or an active call. Sending text does not start a call. If a native Runtime combines that text with work already underway for a voice conversation or a background notification, the combined answer may also be returned through voice.

Text sent through Chet's main input, observed voice transcripts and voice delegations, and their available work receipts and results are retained on the machine running CCCC as Chet's history. Drafts that have not been sent stay in this browser and are shared by tabs using the same Web address; editing the draft concurrently in two tabs can overwrite a change. A receipt means work was accepted, not that it finished; if the outcome cannot be confirmed, Chet says so and does not automatically replay it. Voice transcripts record what this browser received and can be incomplete. There is no history-clear action in this version.

Spoken entries carry a voice icon instead of a text label, and Chet's spoken lines appear as lighter captions without the rule that marks its written results. A written result for a voice request has a quote icon, with an arrow, after its time; it expands the Voice request, Chet's task wording rather than a verbatim quotation of your speech. Adjacent saved Chet voice transcripts from the same call are grouped in an expandable row. The most recent loaded group for an active call starts open unless you change it; other groups show a preview. A folded group is marked if any transcript is unfinished or incomplete. While available, live transcripts and warnings that a transcript was not saved or its saving is unconfirmed appear outside these grouped rows. Voice transcripts and written results arrive independently, so their display order does not establish a one-to-one relationship.

Messages from Groups ​

A Group member can address Chet with @chet as the only recipient, using Send or Request reply; Mail is not available for Chet. In the Web UI, a CCCC administrator can choose @chet in a Group's recipient row. If a member's ID or title is chet, @chet is ambiguous and the send is refused. Chet's messages in a Group are labelled Chet, and a reply to one without other recipients goes back to Chet. Chet writes to a Group directly; it does not forward messages between Groups or send through Connect.

A Group message to Chet appears in Chet's conversation with its sender, Group and time, and Open in Group opens it there. Its text is read from the Group when the conversation loads: a long message is shortened and attachments stay in the Group. If the Group or the message was removed, or the Group was re-created, the entry says the message is no longer available. Group messages that Chet can take up are queued as turns of their own, and Chet is instructed to answer questions and requests in the Group each came from, while an acknowledgement alone needs no reply; your typed or spoken requests go ahead of Group messages still waiting (on OpenCode, only while Chet is answering a Group message; see below), and a Group turn already running is not interrupted. Once Chet has finished, the entry links the reply it finds in the Group, which View reply opens, or says that no reply is seen yet; a reply in the Group does not show that the sender has read it. A Group message whose submission cannot be confirmed is not resent automatically. While the CCCC Web service is stopped, messages to Chet wait in their Groups and are taken up when it starts again; while Chet's Runtime is unavailable, they wait as well. A message whose sender has left the Group or been replaced by the time it is taken up can remain only in its Group. A message in Chet's conversation is checked again just before it is handed to Chet: if its sender has left the Group or been replaced, or the Group was removed or re-created, the entry says it was not handed to Chet.

When Chet runs on OpenCode, a Group message cannot be handed over while messages from your own conversation with Chet are still waiting for Chet to take them up. A Group message that waits only behind work Chet is already doing simply runs after it. If, ten minutes after the Group message was sent, it is still held back by your messages that Chet has not taken up, it is not sent: the message in the Group says it was not sent to Chet and did not run, and can be sent again, and Chet's history says the same; a member who sent it also gets a notice. Your conversation with Chet carries on, and neither notice includes its text or status. Once a Group message is queued with Chet, messages you send to Chet go after it, unless Chet is answering another Group message at the time.

Group messages enter Chet's usual session, which can also hold your own conversation with Chet and other Groups' content. Chet is instructed not to carry that into a Group and to pass requests for work outside a Group on to you rather than do them. These are instructions to the model, not isolation or a permission limit: a model or Runtime can still disregard them.

While Chet's conversation is closed, its entry shows a dot when history holds something new: a Group message, an answer, or a request that did not start. Opening the conversation in this browser clears it. In a browser where the conversation has never been opened, any such Group message, answer or request already in history counts as new.

During a call, a message Chet sends to a Group while the call is recording counts as a call-time handoff, so the member's reply can be read aloud. The recording state approximates this; it does not show which input started Chet's turn. A request you typed during the call can therefore be included, and a handoff Chet sends after the call has ended is not read aloud, although a member's reply to it can still be taken up by Chet when the CCCC Web service and Chet's Runtime are available.

During a call, investigations include a bounded recent excerpt so Chet can understand follow-ups and retain your limits; that excerpt becomes part of its retained conversation. Voice is instructed to prepare a self-contained task and clarify essential ambiguity before handing it off. Handoffs follow the user's current request language and preserve technical identifiers and explicit output-language requirements; an already-clear request can keep its original wording. The collapsible Activity pane shows the selected Chet Runtime's writable terminal, or its structured output for runtimes without one. Work status is independent of live audio. On desktop, drag the divider to adjust the pane widths; focus it and use Left/Right or Home/End for keyboard control. Double-click the divider to restore the default ratio. This browser remembers the chosen ratio, while both panes retain a readable minimum width. Resizing does not reconnect the terminal. On narrow screens, Conversation and Activity switch between the two panes. Settings replace the console body inside the same panel; call controls stay available. Back to conversation or Escape returns to the console without disconnecting its terminal. Chet settings drafts survive tab changes and panel minimization until explicitly saved or discarded. The minimize control keeps the call running; Stop voice ends live audio. Its settings tabs are Call preferences, Message notifications, and Runtime. Microphone and supported speaker choices are shared with Voice Secretary in the global audio section. Call preferences contain the voice service and speaking voice, alongside response detail (concise / standard / detailed) and speaking style (natural / direct / patient). Expression defaults save automatically for the instance and apply to the next call, without restarting Chet. Detail applies to incoming agent notifications as well as answers. Voice is instructed to identify each notification's Group and sender. Detailed asks it to retain substantive findings, numbers, conditions, qualifications and next steps in natural speech; raw tool traces remain in Chet's terminal. For completed investigations, concise and standard ask Voice to prefer a brief spoken account over reading a long formatted result aloud; detailed keeps the requested detail. Your spoken request can override these defaults. These are instructions to the voice provider, so its output may still differ from the saved result. Choose Codex Voice or Grok Voice beside the input on desktop, or in Call preferences. On a phone the choices are in settings. The service and its speaking voice are fixed during a call; end it before changing them. This browser remembers the selected service and each service's voice independently. Microphone and speaker choices stay shared with Voice Secretary. Switching service does not change the independent Chet runtime or its retained conversation. Spoken instructions can override expression defaults.

On the CCCC host, Codex Voice uses codex login; Grok Voice uses a first-party Grok Build subscription login (grok login) and requires grok on CCCC's PATH. Neither path requires an extra API key. The console distinguishes a missing Grok executable from a missing login and offers a status refresh. Status reads do not start a CLI or refresh login; Start uses Grok's native authentication manager, including its credential lock, and gives the browser only a short-lived connection secret. Grok uses grok-voice-3.0, direct browser WebSocket audio and server voice-activity detection. The browser resamples its microphone to 24 kHz PCM and plays output through the shared selected speaker. An interrupted response stops locally at once. An investigation tool acknowledges reception after CCCC accepts it; the actual result returns later through the same global Chet.

The host advertises each service's supported voices and default; Web has no second voice allowlist. Codex offers 25 voices, including Himari, Miko, Rindo and Rivage, with Cove as its default. Grok currently offers 28 built-in voices with Eve as its default. The menu shows documented primary-language hints for new Codex voices and supports searching by language; these hints do not restrict the conversation language. A saved choice is preserved until the host catalog confirms it is unavailable. Account eligibility, service availability and subscription limits remain provider-owned; CCCC does not promise free or unlimited voice usage.

Notifications are off for every Group by default. In Message notifications, choose To user or All chat messages for the Groups you want to hear from; enabling a scope starts at the current ledger boundary, not old history. Explicit replies to work requested through Voice are followed independently of subscriptions. Both acknowledgements and later replies are retained, and only the expected agent's reply to the exact source request counts as its response.

Do not repeat messages already viewed in Web is on by default. It uses exact source IDs after at least 1.5 seconds of stable, fully visible, expanded display in a focused foreground page. Hidden, covered, folded, partially visible and virtual-list overscan rows do not count. This does not change agent Mail unread state or reply obligations. A viewed explicit reply still updates Chet's context, but does not trigger a duplicate reading. A user's explicit request to read or explain a message still takes priority.

The console's compact Message sources section opens the original chat without stopping voice. Pending and unconfirmed counts describe delivery observations, not task completion or proof that you heard a result. During your call, queued output is shown as waiting for Voice delivery; expand the section to see whether it is waiting for the connection or a notification check. Updates continue reaching Voice during conversation; Voice handles speaking and interruptions. An unfinished earlier turn cannot block later delivery. If the connection stops accepting updates, the call stops with an explanation; restarting voice can deliver known unsent notifications while Chet retains its context. Each source also shows the associated result's delivery state. Fully skipped output identifies whether it was already viewed, excluded by policy, or unavailable at the submission check. The recorded skip reason survives later preference changes. “Submitted” describes delivery to Voice, not completed playback or proof of hearing. The experimental provider can omit details or fail to resume a report after an interruption even after receiving its context; Chet and source links remain available to inspect it. Delivery uncertainty can remain for earlier updates while new ones proceed. If notification consumption pauses, a call-wide notice explains that source messages are retained and starting a new call retries consumption; voice and Chet state remain separate. If a result combines viewed and unviewed background messages, Voice can point to the remaining sources rather than inventing a separable summary. Attachments are identified without opening or uploading their contents. Notification text is source data, not authorization for an agent action.

Delivery references and cursors survive restarts in protected instance state. Unsent browser output can return on a later call only when it is positively known not to have been submitted; an uncertain handoff is retained and is not automatically replayed. Available results are returned in their recorded completion order, including catch-up on a later call. All new notification routes require trusted-local/admin access and are unavailable to scoped users and Exhibit viewers. Remote administrator tokens are rechecked during calls, notification handoff and Chet's terminal input/output. Revoking or downgrading one closes even an idle terminal connection within the next authorization poll (one second) and stops new Voice work on that connection. An unfinished update from a replaced Chet session remains unconfirmed rather than being silently executed again by its replacement.

The text and voice interfaces use the same assistant:

  • Realtime Voice owns the low-latency spoken conversation, interruption, and clarification.
  • Chet uses a managed Runtime session for file inspection, tools, current CCCC facts, and substantial analysis. Voice delegations can return progress and results to the spoken conversation. Under local user authority the Runtime has tools to list Groups and query an explicitly named Group; this first slice does not yet provide the planned cross-Group summary and hand-off receipts. See Runtimes for supported modes and their validation limits.

CCCC starts Chet in one stable neutral directory at CCCC_HOME/state/codex_voice/analyst-workdir/. It is not a repository, Working Group, implicit MCP target, or authorization boundary. Voice therefore starts without a selected Group or attached repository and never changes identity when the sidebar selection changes. Chet's CCCC MCP session has global user authority and the full tool catalog, but it receives no default group_id; every Group, agent, task, ledger, or repository operation must resolve and pass an explicit target. Repository modification remains work for the target Group's Foreman or peer rather than work rooted in the neutral Voice directory.

An embedding application can provide optional application_context in POST /api/v1/codex_voice/calls: { "id": "work:123", "instructions": "Reply in Japanese for the current work." }. The ID is 1–128 ASCII letters, digits, -, _, ., or :; instructions are nonempty UTF-8 text up to 24 KiB (24,576 bytes), without control characters except newline and tab. The same byte limit applies to assistant and persona modes. CCCC rejects oversized text instead of truncating it. This is a local input size limit, not a provider token limit; token counts vary with the text and model, and provider limits still apply. CCCC holds this context unchanged for that call. In assistant mode, it includes the context in Realtime startup instructions and every assistant delegation, and uses a context-aware greeting. Replaying the same client session and SDP with different context returns busy rather than silently reusing the old call. Omitting the field preserves the global Voice behavior. Context is not authentication, a selected CCCC Group, a tool permission, or an isolated Chet session. The host application remains responsible for authorization, business records and any session reset needed when changing subjects. Do not include credentials.

For host-defined roleplay, set application_context.mode to "persona":

json
{
  "client_session_id": "training-call-1",
  "offer_sdp": "<WebRTC offer>",
  "voice": "cove",
  "application_context": {
    "id": "training:customer",
    "mode": "persona",
    "instructions": "Act as a customer in a Japanese phone-training exercise. Wait for the trainee to speak first."
  }
}

The default mode is "assistant"; explicitly specifying it behaves like omitting it. In persona mode CCCC supplies only the host's instructions, without its assistant role, routing instructions or saved expression preferences. It does not resolve, launch, attach or reset Chet's Runtime. Incoming delegation events are ignored; agent/Group notifications are neither consumed nor spoken. An existing Chet session and its pending notifications remain available for later assistant calls. Authorization checks, the single-call microphone lease, heartbeats and disconnect cleanup still apply. The host decides the opening behavior in its instructions; CCCC sends only a neutral call-start cue.

Check GET /api/v1/codex_voice/calls/active first: readiness.supported_modes advertises ["assistant", "persona"]. For persona, check realtime_credentials_available; do not require analyst_runtime_available or configure/reset Chet. Credential presence is not a guarantee of provider availability. Old servers reject the nested mode field; never retry a rejected persona request by silently dropping it. Start and active responses include call.mode; persona has call.analyst_generation: null and the start response has analyst: null. The active endpoint's top-level analyst still describes the independently managed global Chet session, if one exists. A different mode with the same client session ID returns busy, just like a changed context or SDP. Stop the old call before changing modes.

Experimental provider boundary: persona startup omits the quicksilver delegation configuration and CCCC enforces no local delegation execution. Omission must not be treated as proof that the upstream model cannot emit delegation events. Public GPT-Live documentation describes delegation: null as client mode, not as a disabling setting; CCCC's experimental quicksilver v2 behavior and both opening directions require real-call acceptance before production use. The host instructions do not override the provider's own system rules.

CCCC reads the existing Codex credential only in the native process that creates the provider call; the browser receives the WebRTC answer and bounded session events, not the credential. Use the console to mute the microphone, resume browser-blocked playback, or stop the call. Chet uses the configured Runtime under the same local user authority as CCCC. Its authentication and model provider are independent from the Realtime login: the default launch inherits the normal Codex configuration, while Custom or Runtime Profile settings can select Codex, Claude Code, Grok, OpenCode, Kilo, or official ACP for Antigravity, Copilot, Devin and Cursor, and configure that runtime's provider home, model provider, or API key without changing the Realtime credential path. The two sides exchange only delegations, progress, and results through the CCCC controller. The Runtime's actual permission behavior follows its selected command and mode.

For a pure ACP Runtime, Activity shows structured output instead of a terminal. Chet's main input remains available; busy ACP work uses a visible FIFO. Cancel stops the current task and queued inputs, including while provider admission is pending. When the current task is your own request, Group messages waiting for Chet stay queued; when it is a Group message, Cancel clears everything queued. While this page owns an assistant-mode call, Activity also offers a task input whose settled result is returned to that exact call. Ending the original call does not retarget its queued or late results to a replacement call. If the call ends with text still in that input, the draft stays visible. Tasks entered through this separate input appear in Activity, not in Chet's persistent conversation history. Activity pairs task text with queued, working, completed, failed or cancelled outcomes, retaining the eight most recent settled rows and any unsettled/awaiting-output work for the current Runtime generation. Enter adds a newline; Ctrl/Command+Enter submits. Failed submissions preserve the draft and original call scope; editing starts a new explicit submission. Returning context to Realtime is not proof that the result was spoken. Independent native-terminal tasks do not automatically become voice requests. Chet's Runtime settings use the same Custom / Runtime Profile model as agent editing. Custom mode accepts a command for any supported runtime (including Kilo's official Windows npm entrypoint), supported provider/model options, and write-only private environment values, and it can save that complete configuration as a reusable Runtime Profile without revealing secrets to Web. Runtime Profile mode resolves a compatible Profile's command and private environment from the shared profile store; the agent-only submit field does not alter the Voice host. Historical Profiles with a string command are accepted and normalized to the canonical argument array when next saved.

Antigravity is supported only in official ACP mode and shares the agent adapter and separate CCCC-managed Google login. It shows structured progress, results, queued input and one-time permission requests instead of a terminal. An existing Antigravity TUI Profile is not eligible. Run cccc setup --runtime antigravity --runtime-mode acp --login once before using it. The Voice console gives this command when the ACP component is not installed; native agy availability and its login do not satisfy this prerequisite. Reading status does not download the component or start authentication. An explicit --model is reapplied on resume; --dangerously-skip-permissions selects unrestricted operation, otherwise permissions require Allow once / Deny. ACP serializes additional inputs in a visible queue of up to 32; cancellation clears current and queued inputs, except Group messages waiting behind your own request. Other Runtime input behavior and Realtime credentials stay unchanged.

For Codex, CCCC removes conflicting host flags and pins app-server, shell_environment_policy.inherit=all, approval_policy=never, sandbox_mode=danger-full-access, the neutral Voice workspace, and the global CCCC MCP binding. For Grok, CCCC accepts a direct root command with supported global provider/model options, then owns the private leader socket, ACP controller, native TUI attachment, YOLO policy, workspace, and per-session MCP binding. Grok subcommands, wrappers, prompt tails, and user-owned topology/session flags are rejected rather than routed through a second PTY path. OpenCode V2 uses a private serve --stdio HTTP/SSE backend with its native --server --session TUI. Kilo uses its own ACP process and kilo attach. CCCC owns the session, authenticated loopback backend, one-time permissions and scoped MCP injection. Supported model, agent and logging choices remain configurable; pure mode is Kilo-only. Kilo's official Windows npm entrypoint uses Node and the installed launcher. Custom wrappers, subcommands, prompt tails, and user-owned topology or session flags fail explicitly. OpenCode requires V2 2.1.4 or newer; old V1 receipts and history are retained for explicit migration or a deliberate New session. These CCCC-owned values cannot be overridden.

For Claude Code, CCCC requires both the launcher and background worker to report version 2.1.259 or newer. Their versions can differ after a Claude update; an upgraded supervisor may retain older workers or migrate idle sessions without changing the managed session's identity. CCCC owns the Agent View background session, authenticated control channel, transcript lifecycle, native claude attach, session identity, MCP binding, autonomy, and resume arguments. Supported model, effort, agent, tool, plugin, and settings options remain configurable. Private Profile environment is merged into a protected CCCC settings file so the background worker receives it without exposing raw values in the job record or Web API. Claude wrappers, renamed binaries, prompt tails, print mode, remote-control mode, and user-owned background/session arguments fail explicitly; there is no Hook, PTY-paste, or claude -p fallback.

For runtimes with a native terminal, the controller and TUI are separate processes in one configured Chet session. Codex's app-server and remote TUI receive the same executable, model, Profile, supported -c overrides, YOLO policy, and private environment. Claude's Agent View controller, transcript follower, and native TUI share one background session and one effective launch identity. Grok's leader, ACP client, and TUI share one resolved provider configuration, private environment, and exact session; CCCC adds topology arguments only to the processes that accept them. OpenCode V2 uses its HTTP/SSE controller; Kilo uses ACP. Both attach a native TUI to one authenticated provider backend and exact session. CCCC injects the scoped MCP binding when it creates the managed session, so the model reached from either client has the same CCCC tools. Opening the embedded terminal therefore observes and controls the existing Chet session instead of starting another model conversation.

If Realtime Voice reports a provider error, its error code is shown separately from Chet failures. Browser and server diagnostics retain only bounded code/type/event/parameter identifiers and the call generation. Provider explanations can quote conversation content or credentials and are not logged. A stopped audio call does not discard the warm Chet session. Provider diagnostics do not retry requests or change the existing disconnect policy.

For startup or later connection failures, keep the CCCC and Runtime versions alongside the first server-log diagnostic. Realtime startup reports its stage, elapsed time, and available request/TLS categories and OS error code. Managed Codex disconnects report the session generation, child-process state and WebSocket protocol category. reset_without_close_handshake means the local transport ended without a WebSocket close handshake; it does not by itself identify which process or network component caused the closure. A later analyst_disconnected Voice-control message can be a consequence of that first failure. A still-running child PID does not prove that its session is usable. These reports omit raw errors, credentials and conversation content.

Ordinary Codex, Claude Code, Grok, OpenCode, and Kilo agents use the same runtime-specific managed adapter as Chet and always attach the Runtime's native writable TUI. Agent controllers are bound to a concrete Group and agent MCP identity, while Chet uses the global user identity and its neutral workspace. These are separate roles over the same adapter, not divergent session paths. Agent messages enter the native TUI immediately, leaving queue-versus-steer behavior to the receiving Runtime. Realtime Voice decides whether a spoken request needs Chet, but it does not schedule Chet's process. CCCC attempts each correlated delegation immediately: Codex can append an explicit in-flight correction through exact-turn steering, while the other admitted Runtimes receive the exact input through their verified native terminal and own whether it steers or queues. Busy state alone never drops or delays a delegation for these runtimes. Antigravity ACP has no native busy-input surface, so its additional inputs instead enter the visible bounded queue described above. All admitted Runtime commands fail explicitly when they cannot join the managed session instead of falling back to a second execution path.

Custom non-secret settings live in settings.yaml; custom private values live in CCCC_HOME/state/secrets/codex_voice_analyst.json. Profile secrets remain owned by the Runtime Profile store, and Web receives key names only. Applying a valid setting while idle restarts the managed host and resumes the same materialized Chet session. Runtime settings cannot change while Realtime Voice is connected. If Chet still has active or queued work after voice stops, Web shows that distinct state and asks once before stopping the old managed host, discarding the unfinished result, and applying the new settings. An effective Runtime Profile command/environment change replaces the warm host at the next voice start while retaining the same session when its runtime identity is unchanged; name, submit, and capability-only edits do not restart this host. Changing the selected runtime, any effective Claude launch setting or private environment, CODEX_HOME, GROK_HOME, OpenCode's storage/config roots (including OPENCODE_DB), Kilo's storage/config roots (including KILO_DB), HOME, or USERPROFILE changes that identity and starts a new session after explicit confirmation in Custom mode. A failed candidate launch restores the prior settings and runtime, but work explicitly discarded before the switch cannot be recovered. Browser-local audio choices apply to the next call.

User requests and source-message updates enter the same Runtime admission path immediately; the Runtime owns steer/queue behavior. Results intended for Realtime are delivered in order, including while someone is speaking. Realtime owns speech timing and yielding to interruptions; missing speech-turn events cannot block later result delivery. Adjacent fragments of the same result are combined without losing text and split to the provider limit before sending. Delivery waits only for connection readiness, send-buffer capacity, and pre-submission policy checks. If the connection or send buffer does not recover within 15 seconds, the call stops with an explanation and positively unsent results remain available for a later call. Previously submitted output is not automatically replayed. Pre-submission checks have a 10-second deadline and retry transient connection failures up to three times with increasing delays. A failed check never permits unchecked content to be spoken. If checking still fails, the call stops with an explanation and positively unsent results remain available for a later call. A successfully suppressed result stays suppressed if its check is retried. If Realtime does not confirm receipt within 30 seconds, the console shows a notice without disconnecting or replaying the update. Results exceeding the 32 KiB Voice limit are reported explicitly: use Chet's terminal for the full output or ask for a shorter summary. Neither notice stops Chet. Provider receipts confirm receipt, not that every detail was spoken.

Realtime Call and the Chet Runtime have separate lifecycles. Stopping voice disconnects browser audio and releases the microphone lease, but keeps Chet warm. Ongoing Chet or linked Agent work continues. Completed Chet results are retained without waking audio. Agent replies arriving while voice is off stay in the Group's chat and ledger, and do not start notification model work in the background. Chet looks for correlated replies only during an assistant call, and handling them is not guaranteed. A later call reuses the warm Chet session. If the selected Runtime supports a native TUI, Activity embeds it for that same session; otherwise Activity shows structured output. It stays available after the call stops and does not create an agent or a second session. Codex creates its resumable rollout lazily, so its terminal appears when the first real investigation starts; Claude, Grok, OpenCode, and Kilo managed TUIs can attach as soon as their sessions are ready. CCCC does not create a token-consuming placeholder task. New session is the explicit reset action and is available only after the call stops. Raw local paths, provider session IDs, loopback endpoints, and external terminal commands are deliberately not part of the product surface.

The embedded native TUI remains writable while managed work is active. The receiving Runtime owns whether new terminal input steers the current turn or queues for the next one; CCCC registers exact Voice input before writing it and uses the Runtime's authoritative echo to associate the result with the correct delegation. OpenCode V2 Chet results follow consumed inbox IDs, model output and terminal execution events on the same ordered backend stream. Prompt admission alone is not completion. Exact-turn interruption is unavailable in V2's public API; use the native terminal to interrupt, or stop the owning session. A completed Voice turn with no result is reported as failed instead of silently returning an empty answer.

The session manager keeps at most one voice call and one warm Chet Runtime per interactive Web host. The normal launcher permits one such Web process per CCCC_HOME, while the call also uses the existing daemon-owned global recording lease shared with Voice Secretary. Hiding the details panel, switching Groups, or leaving the tab in the background does not stop or retarget the call. Application-level heartbeats keep the otherwise idle event WebSocket alive through ordinary proxies. Explicit cross-Group CCCC queries do not retarget the neutral Chet workspace. Use Stop voice to disconnect explicitly. Closing or reloading the owning page or losing its browser transport ends only the live call and releases the recording lease. Stopping Web also stops the warm Chet process and embedded TUI while retaining only the exact materialized neutral-workspace thread receipt in global runtime state. The next call attempts to resume that thread; a legacy repository-bound receipt is deliberately replaced with a fresh neutral-workspace thread and a visible migration warning. A resume failure likewise starts fresh with a visible warning rather than silently changing scope. This Experimental surface does not automatically reconnect a browser audio call. Current live evidence covers Linux x86_64/WSL2 and isolated Chrome desktop plus 390px responsive journeys; macOS, native Windows, mobile browsers, and physical-device audio quality remain unclaimed.

Embedded browser views ​

ChatGPT Web Model, NotebookLM sign-in, and Presentation use the same embedded-browser viewer. The website always runs in the daemon-owned browser session; changing the viewer does not replace that browser, navigate it, or change its profile.

  • Page shows the website content directly and uses the available panel space efficiently. It is the default for normal Web Model operation and Presentation.
  • Browser shows the complete browser window when a safe VNC projection is available. It is the default for sign-in and setup surfaces where browser UI or native prompts may matter.

Switching views reconnects only the viewer transport and keeps the current browser session and URL. On platforms or installations without the VNC capability, Browser is unavailable and Page remains active. Neither view emulates the website: it still sees the same daemon-owned browser process. Web Model and NotebookLM use a real system Chrome/Edge session for sites such as ChatGPT and Google; Presentation may use its own Chromium runtime. Browser-native UI that is outside the web page is only visible through Browser (or through the physical browser window on platforms that expose it).

In Web Model settings, Open ChatGPT reuses the current page without reloading it. Complete sign-in and any website security verification manually in that browser. Browser connected describes the viewer connection; it does not mean ChatGPT is signed in. Delivery waits for the signed-in conversation composer, leaving the sign-in or verification page in place. Use Reload ChatGPT page only when you intend to restart the dedicated browser; its saved profile and delivery target are retained.

Performance behavior ​

  • Hidden tabs release their group event streams immediately. Returning to the tab reconnects and catches up from the last event cursor, so several open tabs do not exhaust the browser's per-origin connection pool.
  • A visible tab keeps its streams. When its window regains focus or the tab comes back into view, the page checks whether the stream has gone quiet or fallen behind the server's latest event. If so, it starts reconnecting, and a reconnect waiting out its delay starts immediately; a successful catch-up can retrieve missed messages.
  • Hover-prefetched group data is reused when that group is selected; group bootstrap does not issue a second agents read for the same transition.
  • Production Web responses negotiate Brotli or gzip compression. Voice Secretary code and locale resources are loaded on demand instead of being part of every initial page load.

Managing Groups ​

Creating and editing a Group ​

Click + New in the sidebar, select the workspace directory and create the Group. You can also run cccc attach /path/to/project on the CCCC host. Select a Group in the sidebar to open it; use the pencil beside its header name to edit its title and description.

Group settings ​

Open Settings and more, then settings, and select This group. Guidance, automation, delivery, messaging, assistants, connections and other Group options remain separate from This instance settings.

Managing Agents ​

Use the + in the Agent bar near the composer to add an Agent. Choose an installed Runtime, set its agent ID and review the configuration before adding it. The Runtime selector reports detection on the machine running CCCC, rather than the browser device. Use Check again after installing a Runtime; a failed check is shown separately from a Runtime that was not detected. Detection does not confirm provider login or successful task execution. Antigravity reports its native TUI CLI and separately installed official ACP component independently. Web Models need no local CLI. Runtime Profiles can still be saved for a different host even when their Runtime is not detected on this one. Open an existing Agent to inspect its Runtime and use its available lifecycle actions. The Group status button controls the whole Group's enabled Agents and message delivery; pausing delivery does not stop work already in progress.

DeepSeek has a DeepSeek connection form in its Runtime settings. Enter an API base URL, the exact model ID supplied by the service, and its API key, then use Save & Restart. Without a saved value the Agent uses DEEPSEEK_BASE_URL and DEEPSEEK_MODEL from the environment it starts with (its own environment settings first, then the daemon's), and otherwise https://api.deepseek.com and deepseek-flash; the model field also suggests deepseek-v4-pro, and compatible gateways may use their own model IDs. A URL ending in /chat/completions is converted to its base URL. The same fields are available when creating an Agent, and saving it as a Runtime Profile includes these values. All three fields use the existing write-only private environment store: saved values are marked Configured, blank fields preserve them, and a new value replaces the saved one. Use default removes the Agent's saved address or model; Remove this key removes a saved key. The launch command remains available under advanced details. Restarting from the Agent inspector does not commit an open configuration draft. A missing DeepSeek credential, or an authentication failure CCCC recognizes, pauses automatic delivery until an explicit launch or restart; pending messages are tried again after a successful start. Custom gateways must support DeepSeek streaming Chat Completions and tool calls. DeepSeek's URL, model, and key explanations are under Connection help, collapsed by default.

Switch to Terminals for the paged multi-Agent view, or open one Agent's inspector from its entry. Runtime restart and session-reset actions have different semantics; use the action's explanation before discarding a session. Use the Runtime terminal's native scrolling to review earlier output.

Claude workspace trust ​

On first launch, managed Claude may ask you to trust its workspace or acknowledge its bypass-permissions disclaimer. The existing agent terminal stays available, and agent details show a persistent Awaiting your confirmation in the terminal notice. Make the choice yourself in that terminal and keep it open. CCCC does not answer the prompt or grant approval for you.

CCCC treats changes to Claude's confirmation files as a reason to retry the managed Agent View launch. It reconnects and resumes message delivery only after that retry succeeds; the waiting notice clears when the updated runtime state arrives. If you decline or exit the prompt, the agent stays stopped for the current daemon run. Use Launch or Relaunch to retry explicitly. Stopping the Group and then starting it again also retries its stopped agents; scheduled wake-ups do not reopen a declined confirmation during that daemon run. Restarting the daemon can attempt to restore agents in a running Group again; it does not grant confirmation on your behalf.

Stopping, removing, or restarting the agent cancels the pending recovery; an in-flight launch is cleaned up instead of attaching after stop. Group shutdown also includes pending confirmation prompts that do not yet have a managed session.

Trust-record monitoring uses the configured Claude directory and the home directory, falling back from HOME to USERPROFILE on Windows in both agent environment overrides and the inherited environment.

Messaging ​

Sending Messages ​

  1. Type in the message input at the bottom
  2. Press Ctrl+Enter / Cmd+Enter, or click Send

On desktop, the input grows with its contents up to a compact height by default. Drag the separator above the composer to set its actual height, including for an empty draft. That manual height stays in place while typing or clearing the input. Double-click or press Enter on the focused separator to restore automatic sizing. Up/Down adjust by 16 base pixels; Home selects the minimum and End the maximum. The manual preference is stored locally in base-font pixels and scales with text size. Its effective height is recalculated as the panel shrinks, capped at 60% of the panel while reserving message space and accounting for other composer rows. The separator is absent below 768 px; mobile message scrolling and the existing input expansion control remain unchanged.

On phones narrower than 640 px, the Voice Secretary sheet keeps language, microphone and device refresh on one row. Prompt-mode help starts collapsed; the Activity area uses the remaining height and scrolls independently, so the last reply can be read without scrolling the control header. Instruction mode places the compact text input and Send button side by side. Document mode stacks the heading, view switch, metadata and actions at full width; paths wrap, markdown headings use smaller phone sizes, and document content scrolls within the available space. Wider layouts use the existing workspace arrangement.

Recipient selections are remembered separately for each Group while the Web page remains open. Sending a normal message keeps the selection; switching Groups restores that Group's selection and unsent draft. Use Clear recipients to return to the Group's default routing. Reply recipients are temporary: canceling or sending a reply restores the normal selection. A temporary cross-group destination does not replace the local Group's remembered recipients; after sending, routing returns to the local Group. Reloading the page clears this in-memory state.

If a selected recipient is no longer available, its chip says so and can be removed. Adding another ordinary recipient does not silently remove it; resolve the unavailable target before sending the message. A reply to a former Group member keeps that member as an unavailable recipient rather than silently choosing someone else.

Broadcasts (@all and @peers, including the Group's default broadcast target) leave disabled agents stopped and deliver to enabled recipients. To wake a disabled agent, address it explicitly; @foreman also wakes the coordinator. An explicit target still wakes when combined with a broadcast. Mail leaves disabled agents stopped.

Messages larger than 64 KiB after UTF-8 encoding are sent as UTF-8 text attachments for same-group and CCCC Connect reply targets. Local cross-group text remains inline because its two local ledgers cannot share one attachment path; the bounded daemon IPC limit covers that JSON route. This applies to typed, pasted, dictated, suggested, and restored drafts. Slash commands still require inline text and therefore reject an automatically attached oversized body.

With an empty message input, press Up to recall your most recent message in the current Group. Continue with Up / Down to browse the already loaded message history. Editing or repositioning the cursor leaves history mode. Recall restores message text only; recipients, reply context, attachments, and delivery options remain those currently shown in the composer.

Message attachments ​

Click an image, a video's first frame, or a previewable file name. It opens in the desktop resource panel while that panel is visible. When the panel is collapsed or you are on a phone, the attachment opens in a full-screen preview. View full screen on an attachment tab opens a new preview, starting a PDF at its first page or a video at the beginning. Close the preview to return to the original tab and its reading or playback position.

Use the previous and next buttons to browse the previewable attachments in the same message in order. On a wide screen the buttons flank the preview. On a narrow screen they sit beneath it. The left and right arrow keys also work when focus is on the preview's controls. Within a PDF or HTML frame, use the viewer's own controls. On desktop, Open in a resource tab opens the attachment shown in the preview in a resource tab.

Video frames load when their message comes into view. Click a video to start playback, or press Play after reaching it with the previous and next buttons. The player uses native browser controls and pauses other message videos. A video also pauses when its tab is no longer in front. Show in messages in the panel returns to the message. Downloads remain available when the browser cannot decode a video. Confirmed uploads play and download from the Group's stored blob, so they remain available after the local upload preview expires.

Markdown, HTML, PDF and common text files have a download button beside their name. Markdown renders by default. View source and Preview switch between its unchanged text and rendered view without downloading it again. Text and logs stay plain. HTML uses the same controls and opens as a static document: scripts, forms and external resources are disabled; inline styles and embedded images are supported. Links can be opened explicitly. An attachment has no directory for relative images, stylesheets or sibling files; it is not a running website or application preview. Download the original when you need those. A long Markdown or text file shows its first 1 MiB, and so does the source view of a long HTML file. HTML up to 20 MiB renders in full; larger HTML shows its first 1 MiB as source only. As in Files, wide slides stacked in a centered column start at the left and scroll sideways. PDF uses the browser's own viewer, or CCCC's reader on touch devices and in browsers without one, as for workspace files. Where neither can show it, download it or, unless CCCC is embedded through Connect, open it in a new tab. Other files stay downloads.

Audio files keep their download chip and gain a play button. A message has one player, which appears below its files; choosing another file switches to it, so you can compare several clips. Starting audio pauses other message audio and video, and starting a video pauses audio. Playback stops if the message scrolls far out of view or you change Group. Which formats play depends on the browser. A file that fails to play shows a note; you can try it again, and its download remains available.

An address that starts with http://, https:// or www. is a link. Links to an image, audio or video file open a page-level preview and load only when you click them. Whether they show depends on the source and the browser, so the preview offers a way to open the original in a new tab. Markdown, static HTML and text from raw.githubusercontent.com or gist.githubusercontent.com also open there, as do full links to this site's Group files, with a download. Other links, including a PDF on another site, open in a new tab. Ctrl, ⌘, Shift and the middle button keep their usual behavior.

A path with a directory, such as docs/guide/web-ui.md, out/全片.mp4 or src/main.rs:12, opens a supported file in the resource panel, as it is now rather than when the message was sent, and so does a Markdown link to a file in the workspace. Names may use letters and digits of any language. Put a path in inline code when a name has spaces or symbols other than ., - and _, or when a relative path starts with a folder named in letters other than A to Z; otherwise it may be missed or only partly marked. Paths outside the Group's active scope stay plain text, and so do paths in a relayed message or a message from another scope, where the same path would mean a different file.

Message diagrams ​

In Chat and Inbox messages, a fenced code block labeled mermaid is rendered as a diagram after the message is complete. Markdown opened in the Files preview or shown in Presentation renders these blocks the same way. Use View source to inspect or copy the original definition. Invalid or oversized diagrams fall back to the source automatically; other Markdown surfaces continue to show Mermaid fences as ordinary code blocks. Flowchart image shapes (@{ img: ... }) also remain as source because Mermaid waits on browser image decoding before completing the diagram and can otherwise block later message diagrams. Click a rendered diagram, or use Expand, to open a near-fullscreen viewer. The viewer reuses the completed SVG without rendering the diagram again; small diagrams expand to the available canvas while large diagrams remain scrollable.

@Mentions ​

Type @ for an Actor or # for a Group at the cursor, including in the middle of a message. Choosing a suggestion keeps the text after the cursor. Chosen references are highlighted in the composer. Displayed messages highlight references to the Group's current agents and selectors and to the Groups a message refers to; a removed agent or an old name may no longer be recognized. A reference to an agent takes the colour of that agent's name in message headers, a Group the link colour, and selectors such as @all stay neutral. Links, common email addresses, and code keep their normal formatting. Message-body references provide context rather than changing the recipient controls. A # Group reference is a delegation hint, not a direct cross-Group send.

Available Actor references include:

  • @all - Reference all agents; reserve broadcasts for announcements or urgent shared constraints
  • @foreman - Reference the coordinator for planning, routing, or a summary
  • @peers - Reference all peers
  • @<actor_id> - Reference a specific agent for targeted work

For concrete delegated work that needs an owner, done criterion, evidence, handoff, or acceptance trail, use task-backed delegation. In chat it appears as a linked task chip; ordinary messages remain the right path for quick questions and discussion.

Replying ​

Choose Reply on a message to quote it and reply in the composer.

Quick reply, in the More menu of a message from one of the Group's agents, sends a ready reply at once without the composer, under the caption Quick reply · sends immediately. It goes to that agent alone, quotes the message, and a notice confirms it. Chet's messages, messages from or sent to another Group or CCCC instance and questions answered in their own card have no quick replies; reply to them in the composer. When CCCC cannot confirm that a quick reply was sent, the notice says so and offers Retry; check the messages first, because a retry can send the reply twice.

The menu lists your quick replies in the order you saved them. A new list starts with four default replies, Agree, go ahead, Explain why, Try another way and Give me choices, each led by an emoji. Manage the list in Settings › This instance › Quick Replies: add, edit, reorder or delete up to 24 replies, the default ones included, each with a name and a text that is sent exactly as written, then Save. A default reply's name or text that you leave unchanged follows the interface language; one you change stays as written in every language, and the reply is shown without its emoji. Restore default puts a changed default reply back as it was, and a deleted one can be brought back from the same page: it returns at the end of the list, as its default, without changing the other replies. If the list was saved from another window or device in the meantime, your edits stay on screen and Reload replaces them with the saved list.

Each time you open the menu or the template list, it reads your replies again and shows a change saved on another device once that read succeeds; if the read fails, it keeps what it showed. Before the first read succeeds, neither offers a reply; if that read fails, they say so, and opening them again tries again.

To start from a reply and edit it first, use Insert a template next to Attach file in the composer. Pick one of your quick replies, filtering by name if you like, and its text is inserted at the cursor, replacing any selected text. The list opens above the input, near where you are typing, or across the input on a phone, and each row shows the name and the first non-empty line of the text. Manage templates at the bottom of the list opens the settings page.

The daemon keeps your quick replies for each signed-in user: browsers and phones signed in as the same user see the same replies, and a browser on this computer that is not signed in has its own.

Project Context ​

Open Project Context from the clipboard button in the Group header. The header identifies the Group being viewed.

A single navigation row opens Tasks, AGENTS.md, Agent State, or Skills. On narrow screens, scroll the row horizontally; keyboard users can move focus with Left/Right or Home/End and activate with Enter/Space. Switching sections retains brief and file drafts as well as task filters.

  • Tasks is the default section. The goal, current focus and warnings stay visible above the board, except while editing, when the form holds them. Expand Coordination brief for the project background and constraints, or use Edit directly. The disclosure stays open while editing. While there are tasks to show, Planned, Active and Done stay as columns, empty ones included, so each remains a drop target; when there are none, or none match the search and attention filters, the board says so instead. Dragging a card to a column moves the task there; Restore on an archived card returns it to the status it had before it was archived. A parent card shows how many of its direct, non-archived subtasks are done, even when filters or paging hide some of them. Subtask progress never completes the parent. Each card shows how long ago the task last changed, by this device's clock, and a Planned or Active card without a summary shows its first open checklist item. Done and archived cards no longer show blocking, waiting or handoff marks; reopening the task shows them again. A Done card whose checklist still has open items says how many remain. While a search or filter is active, the board says how many tasks match. The filter buttons count every task, so they leave out their counts while a search or assignee filter narrows the board. While searching, a card whose match is not near the start of its title or summary quotes the matched words from its title, notes, outcome or a handoff the card does not show; the quote can repeat words the card already shows. On phone-width screens, where the columns stack, Active comes first.

    Opening a task shows its editor: title, status, assignee, priority and waiting on first, then the outcome, checklist, notes and subtasks, with Add subtask. Parent, handoff, blocking tasks and task type are under Advanced details. The outcome, notes and checklist items grow with their text, so a long record scrolls with the editor and long items are not cut off. Text taller than 10,000 pixels scrolls inside its own box instead. A checklist item in progress shows as partly ticked, and any other status an agent set is shown beside the item and kept until the item is ticked. Moving a task to Done needs no particular notes; a done task without an outcome only gets a suggestion to add one. Task type requirements describe how complete the record is; they never block a change. Escape closes the editor and keeps Project Context open.

    Saving sends only the fields you changed, and only onto the version you opened. If the task has changed since you opened it, the save is refused and your edits stay in the editor to copy before you reload the latest version; an editor without edits follows such changes by itself. When a save cannot be confirmed, the editor reads the task to tell whether it went through, and saving again never applies it twice. A task that no longer exists keeps your edits in its editor until you close it. If creating a task cannot be confirmed, creating it again unchanged from the same form finishes without making a second one; a closed form or a reloaded page has no such guarantee. Delete is offered only for a task that never left Planned, together with its unexecuted subtasks; for other tasks the editor says why, and archiving is the alternative. A deletion is refused, deleting nothing, if any of it changed after you confirmed.

    Above the board, Tasks | Backlog switches to the project's backlog, and each Group keeps the view you left it on in this browser. The backlog is a Markdown file in the Group's project folder for long-lived topics and ideas; tasks made from a topic track the work on it. It stays an ordinary file: CCCC reads and edits it in place, never exports or commits it, and any editor can change it too. A Group without a project folder has no backlog. Until a file is chosen, the view offers to create BACKLOG.md, which in a Git repository can be listed in the clone's local exclude file so git add -A leaves it out (.gitignore is not changed), or to use an existing file. Tasks work without one.

    ## headings are sections, in your own names and order; only a section named exactly Closed holds finished topics. ### headings are topics, and headings inside a topic use #### or deeper. A topic is a row like a task's, with its linked tasks under the title; open the row to read the topic, or a linked task to open that task. Create task… makes a Planned task from it, or opens the topic's unfinished task if it has one. Edit, Move to… and Close… change the topic; closing asks for a one-line outcome, adds it with the date and moves the topic to Closed, where Reopen… moves it back. The first change here gives a topic a hidden ID comment in its heading, so it can be renamed or moved and stay linked; task IDs are never written into the file.

    A task made from a topic shows the topic's title in its editor. Update the entry… takes the task's outcome to the topic, as a dated paragraph or by closing it. Put back in the backlog adds the outcome to the topic, or saves a task from no topic as a new topic and links them, then asks whether to archive the task. Finishing a task never closes its topic.

    Each change applies only to the file as this page read it. If the file changed elsewhere, nothing is written and your text stays; reload and save again. A draft whose topic itself changed is kept to copy, never written over the change. The backlog is read when Project Context opens and again when shown after a change made through CCCC; edits from another editor appear on Reload. The read-only exhibit page has no backlog.

  • AGENTS.md views and edits the file in the current workspace root, showing its actual path. A missing file can be created explicitly; existing documents are not renamed, merged, or overwritten as part of setup. Each runtime controls instruction discovery and precedence. Saving does not reload running sessions; the optional team notification asks relevant agents to re-read the file. This editor does not represent all instruction files loaded by every agent. Web Model and Dots agents can read files through cccc_file when needed.

  • Agent State shows the latest saved reports, including their timestamps and staleness. Focus, next action and blockers remain visible. Expand working context or recovery cues when needed. A task named in a report does not prove that its Runtime is currently working on it. Viewing repository instructions or records does not keep a task board beneath them.

  • Skills shows self-evolving candidates originating in this Group. Instance capability governance has a wider scope; the two views do not replace each other.

Settings Panel ​

Open settings from Settings and more in the header:

System notifications ​

Use Settings → This instance → Notifications to enable notifications in each browser or installed app where you want them. This is available in the private Web UI, not exhibit mode. Keep the CCCC Web service running; closing the CCCC page does not stop the service on your server. While requests wait for you in a Group, and this browser could show notifications for your account but they are not on yet, its messages page also offers them above the message box: Turn on… opens this settings page, and Not now hides the offer in that browser. No offer appears when the browser cannot use notifications, when they are blocked, when the server cannot send them, or when their state cannot be confirmed.

  1. Open CCCC over HTTPS, or use localhost on the same computer. On iPhone or iPad, first add CCCC to the Home Screen and open it from there; Web Push requires iOS or iPadOS 16.4 or later. See WebKit's setup guidance.
  2. Select Enable notifications and allow notifications when the browser asks. If Finish enabling appears, select it to complete the subscription.
  3. Check that the page reports notifications enabled. To turn them off for this browser, return to the same page and select Disable notifications. This works even when the server cannot be reached: the browser stops receiving them, and the server forgets the subscription once its next notification is refused. Signing out of this browser in Web Access also asks the server to stop notifying it for that account.

On desktop, Chrome and Edge need a browser process to remain running. On Windows, Edge can keep running after the last window closes if background mode is enabled in edge://settings/system; see Microsoft's Edge documentation. Whether Chrome can remain running in the background depends on its version and management settings. For macOS, Safari supports receiving Web Push while Safari is quit on supported systems; see WebKit's platform requirements. Browser settings and operating-system notification permissions also affect delivery.

Notifications are for new messages explicitly addressed to the user. Requests already waiting when notifications are enabled remain in Waiting for you; they are not sent again. Events written while the Web service is stopped are not replayed; a sender restart or handover can also leave a gap. During a temporary push-service failure, up to 10 unanswered requests per browser are retried while the service runs; older requests may be dropped. The push service and device may also delay or discard notifications. Check Waiting for you if a notification does not open its request. On Safari 18.5 or later for Mac, and in iPhone or iPad Home Screen apps on iOS 18.4 or later, opening a notification tries to load CCCC at its message; in a Home Screen app this can reload the open page, and unsent text in it is then lost. Message previews may be visible on the lock screen. The server stores a private signing key and subscription secrets in CCCC_HOME, so keep it accessible only to trusted accounts; on Windows file access depends on the directory's permissions.

If the sender is starting or unavailable, select Try again after the Web service starts and check its logs if the status persists. If the browser blocks notifications, allow them in browser or device settings. A saved subscription that is not connected to the current account can be removed or enabled again from this page.

Copy Groups ​

Use Copy Groups when you need to duplicate, migrate, or back up a working group.

  • Export group copy downloads a zip containing durable CCCC group state: ledger history, agents, memory, attachments, automation, and group settings.
  • The copy package does not include the workspace repository/project files. Copy or clone the workspace separately, then choose the workspace root during import. The folder chooser can also create an empty directory if needed.
  • System credentials, browser sessions, provider auth, and live runtime state are excluded. The package still contains user content such as ledger history, memory, and attachments; treat it as sensitive. Imported agents are stopped and the imported group starts idle.
  • If a group id already exists, import creates a new copy instead of replacing the existing group.
  • Pins are not copied: the copied group starts without any. The pinned messages themselves are in its history.

Automation ​

Open a Group's menu in the sidebar and choose Automation, or open Settings and more, choose All settings, select This group and open Automation.

Each rule shows its last run and its next planned run in your browser's local time, as of when the page opened; saving or reopening the page refreshes them. The last run appears once the rule has run and created its notification, marked when the run came from Run now; runs from before this version are not shown. Instead of a next time, a rule shows why it will not run: a paused or stopped Group runs no automation, and a Group set to idle does not run the built-in stand-up. Resuming a Group restarts the interval and recurring clocks, so the next run can move. A recurring rule saved in another time zone shows that zone, and editing it keeps the zone. A schedule the daily, weekly and monthly patterns cannot express shows as written; it stays as it is until you pick one of those patterns.

Run now sends a reminder rule's notification once, without waiting for its schedule. It runs the saved rule, so save your changes first; it also works on a rule that is off. It does not count toward the maximum number of runs and does not change the schedule, so a scheduled run may still follow soon after. A paused or stopped Group has to be resumed first. The notification may wake its recipients and use model credits. If the page cannot tell whether the run went out, it says so and does not try again: reopen Automation and check the last run before running the rule again. An unchanged last run does not prove that nothing went out.

  • Stand-up: The built-in example is off by default. Its default schedule reminds the foreman at two-hour intervals to review whether work serves the user's current goals and outstanding requests; where waiting, rework, duplication, unclear handoffs or misunderstandings cause friction and whether the team fits the work; whether repeated user concerns have been addressed and adopted improvements have helped; and whether there is a better way to collaborate. For a question that needs a deeper look, it asks the foreman to mark it Review: in a related optimization task, or in a new planned one that authorizes no change, and to suggest a team review rather than run one. It asks for a brief report when the user requests one and otherwise for silence when nothing needs attention. The notification may wake an agent and use model credits; you can edit its schedule, recipients and snippet to fit your team.
  • Rules: Create scheduled reminders with interval / recurring schedule / one-time schedule. Repeating reminders can have an optional maximum number of runs; leave it blank for no limit. The list shows progress and completed rules. Editing the limit keeps the count; removing it resets the count, so adding a new limit later starts from zero.
  • Actions:
    • Send Reminder (normal reminder delivery)
    • Set Group Status (operational, one-time only)
    • Control Agent Runtimes (operational, one-time only)
  • Snippets: Reusable message templates managed alongside rules.
  • One-time behavior: A one-time rule completes and turns off after it fires, and can then be cleared from the completed list. A one-time rule whose time passed while the Group was paused or stopped does not run late, and it may also show as completed without having sent anything.

IM Bridge ​

Configure Telegram, Slack, Discord, Feishu, DingTalk, WeCom, or Mattermost integration.

Group Space ​

Connect the provider account and bind the Group's work and memory notebooks. The Web page shows a connection summary; source ingestion, queries and jobs use agent-facing tools or the CLI. The provider-backed memory notebook is separate from the former local Group memory files; CCCC no longer automatically uploads those files. Automatic two-way synchronization is retired.

For end-to-end setup details, see: Group Space + NotebookLM.

Theme ​

Switch between Light, Dark, or Auto theme. Auto follows the light or dark setting of your system.

Both themes use the same pixel-screen look: light mode sets dark ink on a pale screen, dark mode light ink on a dark screen. Surfaces are separated by thin rules rather than shadows, and keyboard focus shows a dashed outline. Body font under Appearance offers Sans, the default, which sets message text in IBM Plex Sans while message headers, controls and code stay monospace; Mono, which sets most text in monospace; and Pixel, a pixel font. Text size also scales side-panel titles, file paths and settings labels; terminal text keeps its own sizing.

Mobile Usage ​

The Web UI is responsive and works well on mobile:

  • Swipe between tabs
  • Pull down to refresh
  • Tap and hold for context menus
  • Works in mobile browsers (Chrome, Safari)

Remote Access ​

To access from outside your local network:

LAN / Private Network ​

bash
CCCC_WEB_HOST=0.0.0.0 cccc

This keeps localhost access working while also letting other devices on the same network open http://YOUR_LAN_IP:8848/ui/. The native launcher also honors the binding saved in Settings > This instance > Web Access, including the 0.4.35 settings.yaml during migration. Explicit --host / --port flags still take precedence.

If CCCC is running inside WSL2's default NAT networking, this is the exception: 0.0.0.0 only opens the port inside the Linux VM. For true LAN access from other devices, enable WSL mirrored networking or add a Windows netsh interface portproxy rule plus matching firewall allow.

bash
cloudflared tunnel --url http://127.0.0.1:8848

Tailscale ​

bash
CCCC_WEB_HOST=$(tailscale ip -4) cccc

Security ​

Direct browser access through localhost, 127.0.0.1, or ::1 is passwordless and does not create an Access Token. CCCC grants that request an in-memory local administrator principal only when the reconstructed browser-facing origin is loopback. Unsafe writes and WebSockets must also carry the exact same loopback Origin; non-local proxy client addresses are rejected. This local principal is never persisted and is not valid through LAN, Reach, a public URL, or a reverse proxy.

Before exposing the Web UI beyond localhost, first create an Admin Access Token in Settings > This instance > Web Access. With no administrator token, non-local clients receive only the UI shell and health/session guidance; protected APIs and business WebSockets remain locked, while direct loopback access keeps the passwordless local principal described above. Direct local setup no longer requires copying a bootstrap code. Linking your account initializes an administrator Token if needed; completing that flow from the verified localhost Web page also signs that browser in. Enabling hosted Remote Access from that local page covers installations linked before this automatic setup existed. Existing Tokens are preserved. When first configuring through a remote address instead, read web_bootstrap_token inside the host's effective CCCC_HOME (default ~/.cccc) and enter that one-time code. The file is mode 0600 on Unix and is removed once administrator setup completes.

The Web Access panel keeps LAN/public Save, Apply now, and remote-endpoint copying disabled until an Admin Access Token exists. The native daemon and Web boundary enforce the same rule at remote start, apply, and listener boundaries, so direct API calls and stale saved settings cannot bypass the panel. Group-scoped tokens do not satisfy this administrator recovery requirement. Switching back to localhost-only remains available so an incomplete remote setup can be recovered safely.

In Settings > This instance > Web Access, 127.0.0.1 means local-only and 0.0.0.0 means localhost plus your LAN IP on a normal local host. On WSL2 NAT, it still stays inside the VM until Windows networking forwards it outward.

Save stores the target binding. If Web was started by cccc or cccc web, use Apply now in Settings > This instance > Web Access to perform the short supervised restart. If Web is managed by Docker, systemd, or another external supervisor, restart that service instead.

Before an address change restarts the supervised Web service, CCCC attempts to stop its active Voice work and Chet. If their cleanup cannot be confirmed, the current Web remains available and the saved address has not yet taken effect. Check the Voice state and retry Apply now after cleanup. A Voice start can likewise be delayed while resources from a previous Web process are being recovered; retry after cleanup, and if the condition persists, review the process records as described below. If Chet reset fails after cleanup began, refresh the Voice state before deciding whether to try again.

On Unix, CCCC records the processes it starts so that it can clean them up after a crash. When it cannot confirm that processes from an earlier Web or daemon run have ended, Voice does not start, and the Secretary may hold back new tasks and the automatic recovery of unfinished documents. The Voice error then offers Review earlier processes, and the Secretary panel shows Review. The review lists each record and why CCCC is still waiting:

  • Check again retries cleanup for that record only.
  • Accept stops waiting for that record. It does not stop the old work: if that work is still running, it may keep changing files or documents or sending requests that cost money, and remote work may continue after local cleanup.
  • Accept all appears for two or more records that can be accepted. It accepts them one at a time and stops as soon as anything changes; it is not one transaction.

Some records need offline review. Stop CCCC safely, then run cccc daemon process-recovery --owner daemon list (or --owner web), followed by retry --id <id> --version <version> or accept --id <id> --version <version> --accept-risk. While CCCC Web is running, these commands refuse to change Web's records. Do not delete files under CCCC_HOME by hand. On Linux, an agent that CCCC could not confirm after its launch is reviewed with --owner session; see When CCCC cannot confirm an agent.

Use one CCCC_HOME on one host or container at a time. Before moving or copying it elsewhere, make sure the work started from it has ended. On that basis, CCCC releases records from an earlier boot without signalling anything when it can identify that boot; this does not prove that remote requests ended.

On Linux, cleanup prefers the strongest process identity the system offers and keeps records it cannot confirm for review; some systems still need approximate matching, which can misjudge. On macOS, the crash watchdog still judges processes approximately, by process group number and start time to the second: a reused number, or the gap between that check and the signal, could stop another process by mistake. How often this happens is unknown; earlier versions behave the same way. At the next start, CCCC on macOS signals nothing: it releases a record from this version once its process and its process group have both ended, or a record from an earlier version once its process group has ended, and keeps the others for review.

For the default local app flow, prefer restarting from the owning cccc session itself: Ctrl+C to stop the whole app, then run cccc again. That keeps daemon and Web on the same fresh code/runtime.

Start / Stop are only for Tailscale remote access and do not rebind the already-running Web socket.

CCCC keeps the token policy simple:

  • localhost-only: direct loopback browser requests are passwordless and use a non-persistent local administrator principal
  • LAN/private network and public URL/tunnel/reverse proxy: an Admin Access Token is mandatory before exposure

CCCC_WEB_ALLOW_UNAUTHENTICATED=1 is only an unsafe listener override; it never grants API authorization or bypasses first-admin bootstrap. Plain HTTP manual LAN exposure is allowed by default for private-network use, but still requires an Admin Access Token before remote APIs can be used. Prefer an HTTPS reverse proxy, tunnel, or encrypted overlay for untrusted networks.

CCCC adds frame-ancestors 'self', SAMEORIGIN, nosniff, no-referrer, a restrictive permissions policy, and HSTS on HTTPS responses. Supervised CCCC Web processes trust reverse-proxy forwarding headers automatically only while the effective listener is loopback. A supervised LAN/wildcard listener or externally managed reverse proxy must explicitly set CCCC_WEB_TRUST_PROXY_HEADERS=1 and must overwrite—not append—client-supplied Forwarded and X-Forwarded-* headers. Direct public listeners should leave this flag unset.

Enter the Access Token in the Web sign-in form. CCCC validates it through the Authorization header and establishes an HttpOnly, SameSite=Lax session cookie. The cookie has a rolling 30-day lifetime and is refreshed when the Web session is checked, avoiding repeated token entry after mobile browsers reclaim a tab. The temporary header token is removed from browser session storage after the cookie is established. Access tokens are not accepted in ordinary API, SSE, or WebSocket query strings and should never be placed in shared URLs.

For managed remote access, open Settings → This instance → Web Access and link this installation to your CCCC account in the Reach section. Approve the request on the website, then return to the same local panel. If an administrator Access Token already exists, reuse it; otherwise complete the displayed token prerequisite. Choose Turn on remote access explicitly. Account registration only reserves the device address; enabling Reach creates its DNS record and tunnel.

Once enabled, Reach restores automatically when CCCC restarts and its Web listener is ready. Keep CCCC running on this computer; opening the settings panel is not required. Temporary network/account failures retry in the background, with a delay of up to one minute between failed attempts. An already-running tunnel helper handles its own network reconnection. Turning Reach off, unlinking the account, or cutting the device prevents automatic restoration. If unlinking cannot complete because the account service is unavailable, Reach stays off and the device credential is retained so you can retry unlinking. If the installed helper is missing or no longer matches this CCCC version, install it explicitly with cccc reach install; automatic restoration does not download or upgrade executables.

The panel confirms the connection with at most six status checks over 45 seconds. It pauses checks while hidden and never restarts or provisions a tunnel as a retry. If confirmation ends without a connection, check this computer's network and CCCC process, then use Check connection. A stopped helper offers an explicit startup retry in addition to daemon recovery. Turn off remains available, and other provider/binding settings remain locked while Reach is enabled even if the tunnel disconnects.

Tunnel connected means the account service observed a connected tunnel. Open Web from another network to confirm application access; website account login does not sign the browser into this device. The displayed check time is the last status observation, not a continuous availability guarantee.

Reach follows the same credential rule. Its status payload exposes only a tokenless public address. Clicking Open Web or Copy Admin Link asks the local authenticated Rust Web session for a 120-second, one-time exchange code bound to that Reach origin. The public endpoint consumes the code once, establishes the HttpOnly cookie, and redirects to a clean /ui/ URL.

A passwordless localhost administrator reuses an existing administrator Access Token for this exchange; no new long-lived token is created. A remotely signed-in administrator remains bound to their own active token. Revoking that token also invalidates outstanding links backed by it.

Cookie-authenticated Rust Web writes and WebSocket connections require an exact allowed Origin, with a same-origin Referer accepted only as a fallback. This check is independent of CORS and blocks same-site sibling domains from submitting state-changing forms.

Cross-origin browser clients can opt in through CCCC_WEB_CORS_ORIGINS, a comma-separated list of exact origins (scheme, host and port). Named origins support credentials and are also accepted by Cookie write and Cookie-authenticated WebSocket origin checks. Include only trusted client sites.

CCCC_WEB_ALLOW_ANY_ORIGIN=1 instead enables wildcard HTTP CORS without credentials, for clients that explicitly supply an Authorization: Bearer ... access token and omit browser credentials. It does not bypass Cookie write or Cookie-authenticated WebSocket origin checks, create an authenticated principal, or expose the local passwordless principal to other sites. Local passwordless reads with an Origin or Referer must identify the same loopback origin, just as writes do. Browser WebSocket clients using cookies still need a same-origin connection or a named trusted origin. Explicit Bearer-authenticated WebSocket clients do not require an Origin match; token and Group permissions still apply. If both settings are present, wildcard HTTP CORS takes precedence; use only CCCC_WEB_CORS_ORIGINS for cross-origin Cookie sessions. Both settings are off by default and require restarting the Web process to change CORS responses.

Reverse proxy headers ​

When a reverse proxy terminates HTTPS or exposes CCCC under another host, it must overwrite the browser-facing host and protocol headers. These values are used by every browser WebSocket (realtime events, terminal, Voice Secretary, projected browser) and by Cookie-authenticated write protection:

nginx
proxy_http_version 1.1;
proxy_set_header Host $host;
proxy_set_header X-Forwarded-Host $host;
proxy_set_header X-Forwarded-Proto $scheme;
proxy_set_header Upgrade $http_upgrade;
proxy_set_header Connection "upgrade";

Do not pass through client-supplied X-Forwarded-* values. The trusted proxy must overwrite them. CCCC also accepts RFC 7239 Forwarded with host and proto, and handles comma-separated multi-proxy X-Forwarded-* chains by using the first browser-facing value. A mismatch is rejected with csrf_origin_invalid for Cookie-authenticated writes and WebSockets. If a proxy cannot preserve the public host, configure its exact browser origin in CCCC_WEB_CORS_ORIGINS instead of disabling source protection.

A token scoped to selected Groups receives global stream metadata only for those Groups, and the global stream never carries message content. Full event content remains on the per-Group stream and is subject to the same scope check. Administrative capability changes require an Admin token.

Same-origin browser requests carrying Sec-Fetch-Site: same-origin satisfy the Cookie CSRF check even when a reverse proxy rewrites Host or terminates HTTPS. This does not require CCCC_WEB_CORS_ORIGINS or global proxy-header trust. Proxies should preserve this browser-generated header, never manufacture it for cross-origin requests. same-site, cross-site, none, and missing metadata continue through the existing Origin/Referer and configured-origin checks. Authentication and group permissions remain required. See the Fetch Metadata specification.

Browsers send no Fetch Metadata on a WebSocket handshake, so Terminal and stream sockets fall back to comparing the request Origin against the served host. That comparison ignores the scheme, because TLS terminating at a reverse proxy makes the origin server read every https:// page back as http://. Host and port must still match exactly — a different host, port, or subdomain is rejected — and the scheme alone adds nothing, since an attacker able to serve a page from this host would already control the site. Proxies that rewrite Host still need CCCC_WEB_TRUST_PROXY_HEADERS=1 or an explicit CCCC_WEB_CORS_ORIGINS entry.

Shared realtime connection ​

The Web UI uses one /api/v1/events/ws WebSocket per page for global metadata, current-Group ledger events, and headless output. Switching Groups replaces the subscriptions on the existing socket. Background pages release their subscriptions and close the socket when none remain. This avoids occupying the HTTP/1.1 pool with three persistent SSE requests per visible page; ordinary API requests remain available with several workbench windows open.

Each subscription has an ID. Both sides ignore packets from retired IDs. Reconnects resume the ledger from its last delivered event ID and request a headless snapshot from the same tail that supplies subsequent deltas. Existing UI catch-up and coalescing continue to apply. Global events contain only permitted Group metadata. Token and Connect-frame authority are rechecked while connected; each Group subscription also validates Group access before opening its producer.

The server uses a bounded eight-packet output queue and a five-second socket-write deadline. Heartbeats detect broken connections, and closing the socket cancels all its producers. A failed channel retries independently; transport failure reconnects with backoff. The UI does not fall back to HTTP SSE, which would recreate the connection-pool blockage. The existing SSE endpoints remain available for external clients and use the same typed event producers. Deploy the updated UI and backend together to enable the new endpoint.

Released under the Apache-2.0 License.