Desktop messages

Read Agent replies and execution history in the desktop app

If a saved Agent root reports session_migration_required, its Session data needs offline conversion before it can be read. Opening a conversation does not convert or replace stored data. An existing Agent directory with no database or legacy Session partitions has an empty conversation list; reading it does not create a database. Reads of saved conversation records, missing directories, and unreadable storage still report their actual errors.

A Session keeps the Workspace ownership determined at creation. History remains accessible if the original Workspace is unavailable; restore that Workspace before sending again. The composer preserves your draft. Parent and Child work may use different Workspaces, each retaining its own ownership.

Sidebar rows, expand buttons, and navigation icons support keyboard interaction without drawing focus rings.

Each main navigation entry remembers its last page independently. Returning to Chat restores your last selected Agent or Group and conversation. Stable pages remain available after refreshing or reopening Desktop, even when you last visited Works, Tasks, or a Power. Temporary creation pages and drafts are retained only in the current window. If a saved resource has been deleted, Desktop opens an available page in that entry; a missing conversation returns to its Agent or Workspace.

The Chat sidebar mixes Agents and Groups in descending order of their fixed Chat’s saved Session update time. Project conversations and project group chats do not affect this order. Subjects without a fixed Chat appear last. Equal timestamps place Agents before Groups, then sort names within each kind. Startup and refresh read the saved fixed Chat summaries, and conversation updates automatically refresh the order.

Right-click a session row in the Works sidebar, or click its ⋯ button, to open its actions. Agent sessions offer rename, pin or unpin, archive, multiple selection, quote in the current composer, copy path, copy Session reference, and delete. Group sessions offer rename, pin or unpin, archive, multiple selection, and delete. Both entry points share availability rules and confirmation dialogs; quoting is unavailable when there is no receiving composer. The session row stays highlighted while its context menu is open and returns to its previous state when the menu closes. Right-clicking keeps the current conversation selected, and deletion requires confirmation. In multiple-selection mode, use the row's selection control.

Startup and background connections

Desktop uses schema v12 Session storage and checks the storage format and access coordination before reading history. An incompatible format reports session_migration_required; complete the offline migration first. History reads do not convert databases automatically. Identity, format, and permission failures report their actual cause.

Each conversation has one unique, permanent session_id. Archiving preserves that ID. Recreating a main conversation assigns a new ID, so old pins, notifications, and links cannot redirect to it. Archived conversations remain available for read-only access through history.

Desktop loads its window independently, without waiting for the Agent controller to be created or Agents to finish assembling. Settings, Agent catalogs, and Workspace catalogs load separately and appear as they arrive. Saved Session summaries, lists, messages, and configuration use independent read access and do not wait for Agent execution, model catalogs, Workspace connections, or Power initialization. Sending and other execution operations still wait for their required runtime. A failed request shows its cause while other loaded data remains available. Extension Powers load independently in the background. Bot connections do not hold up the window or local data reads. Loading Powers show their initialization state; connection results are available in each Bot Account.

Local core assembly, Agent and Workspace catalogs, and historical Session reads do not depend on Federation login, connectivity, or the remote model catalog. Local services and saved content remain available while signed out or when Federation is unavailable. Model capabilities and context capacity resolve before execution. Temporary catalog failures preserve configured models; execution failures report their cause. Restoring the last Session reads that Session directly, and pinned entries load on demand. Opening an Agent’s main conversation follows its saved main-conversation pointer to the permanent Session ID and reads the existing record without execution resources. If no main conversation exists, Desktop creates one in the Agent’s private main Workspace.

Opening a conversation selects its target immediately and shows loading while its saved summary is read. The Works sidebar highlights the current root conversation consistently in pinned entries and Workspace groups. When viewing a child Session, its root stays highlighted. Messages and configuration then load independently: available messages appear without waiting for configuration, and a late response from an earlier target cannot replace the current view. Archive retries within the current Desktop window reuse their original operation and root record, so recreating a main conversation does not redirect a retry to the new record.

The Works sidebar loads recent conversations when you expand a Workspace. Each Agent’s page appears as it arrives. Collapsing a Workspace discards pending results and releases queued reads; reopening continues from confirmed pages. Each Workspace initially shows its 10 most recent root conversations. Load more conversations reveals up to 10 additional roots per click, using loaded content first and reading further pages when needed until the batch is filled or available sources are exhausted. The button shows loading progress and prevents repeated clicks during reads, then disappears once all conversations are loaded and visible. The main list shows only root conversations, with a numeric tag immediately after the title showing the loaded descendant count; the root of the current conversation and selected root rows remain visible. The visible conversation stays at the same scroll position as later pages arrive. If an Agent fails to load, Retry loading conversations appears beside its name while other Agents can continue loading further pages. A retry reads only that Agent; loaded content, the current conversation, and pinned entries remain available. Once the other sources are exhausted, the failed Agent keeps its retry action. The full total remains unknown until all sources load successfully. Unpinned root conversations show running Sessions first, then follow Session activity time, newest first, with a fixed order for equal timestamps. The list updates its order immediately when execution starts or ends. Agent conversations advance their activity time when a user message arrives or an Agent message finishes, fails, or stops. Streaming keeps the running priority; opening a saved conversation and renaming it preserve its position. Pinned roots keep their chosen order and show the same descendant count tag. Loading further pages fills in the directory using the same ordering.

In Agent messages, “Waiting for child work” shares the animated dots, text styling, line height, and spacing used by other runtime indicators. The sidebar shows a working spinner while a conversation is submitting, executing, or waiting for child work. A main Session keeps this spinner while its detail view shows “Waiting for child work”; waiting for user input shows a response-needed indicator. Refreshing the Desktop page restores working and waiting-for-response indicators for background conversations in the sidebar without opening them. The indicators update when execution finishes or stops. Status changes received during refresh take precedence over the initial read.

The notification service is ready before extensions load. Persisted background work resumes with limited concurrency after initial Power assembly settles. A failed extension definition, integrity check, module import, or initialization affects only that item; other valid Powers continue loading. Its Power page shows the cause. After correcting configuration or reinstalling the extension, click Retry. Ready Powers are not initialized again, and successfully recovered work is not recovered twice.

Local runtime operations wait for the current core startup attempt and report local assembly failures. Sign-in and account switching are independent of core startup. On exit, Desktop stops accepting new resource changes, immediately requests cancellation of pending Power initialization, and waits for accepted operations and resource cleanup before closing local storage. A Power whose cleanup fails retains its instance and error; Retry first retries cleanup and loads it again only after cleanup succeeds. If shutdown cleanup fails, Desktop reports the failure and retains local storage and its ownership lock. Choose Quit again to retry. A third-party extension that ignores cancellation can still extend the wait.

App updates

Automatic updates are on by default in the official macOS Apple Silicon release. Downcity checks for a stable release at startup and once an hour while it runs. It downloads an available update in the background.

When the update is ready, choose Restart now or Later. If you choose Later, Downcity installs the update when you quit normally. The next launch uses the new version.

Go to Settings → General → App updates to see your current version, follow download progress, or retry. If you turn off Automatic updates, you can still Check for updates and Download update manually. Updates that are already downloaded still install when you quit normally.

If your version has no App updates section, download and install the current release once.

Add and manage code extensions

The plus button at the end of the Navbar capsule opens Add extension without switching the current conversation. Enter a repository address or choose a local project directory or downcity.ts / downcity.js entry. Expand Git ref when needed. Load submits preparation and loading to the main process. Closing the window does not cancel an accepted operation; reopen it to inspect the latest result.

Settings → Extensions lists projects in the current data root, including projects added through CLI. Reload, remove, or expand details to fill in the fields requested by the entry. Secret fields use private input; leave them blank to retain existing values. Save changes, then explicitly reload to apply their new version. Failures show their actual error, and unexecuted projects show Pending load.

Powers with work pages enter the Navbar after complete loading. Their Sidebar, Mainview, and configuration slots share the backend identity. The host compiles browser code separately and supplies its React and Power UI SDK instances. Reload remounts the new runtime generation; late calls from old pages cannot enter a replacement instance. Removal withdraws pages and resource access. See extension project management for commands and development guidance.

Automatic sign-in renewal

Desktop and CLI access credentials last 15 minutes and renew before a request when less than one minute remains. Successful renewal rotates the refresh credential and extends the idle deadline by 90 days. Closing the app does not end sign-in: a saved session can renew when used again, until it is revoked or idle-expired.

Desktop, CLI, and an already-running City share renewal results for the same session. Network failures preserve sign-in and allow retry; confirmed authentication failures require sign-in. Model calls and other writes are never automatically replayed after a 401.

Removing an account or signing out invalidates its local shared session and attempts to revoke the entire remote device session. Offline sign-out still clears local credentials; an unavailable service cannot confirm remote revocation. Other independently signed-in devices keep their own sessions. Accounts without a refresh credential need to sign in to obtain a renewable session. Explicit host credentials supplied through DOWNCITY_USER_TOKEN do not renew automatically.

Switching accounts

When you switch accounts in account settings, Desktop validates the target account before making it active. Validation failure preserves the current selection. Profile refreshes do not switch accounts, and a late response cannot restore a removed account. During successive sign-in, switch, removal, or logout operations, only a still-current identity operation can commit its result. Finish or stop running conversations before activating a newly signed-in account, switching, removing, or signing out of an account.

Sign-in, Federation or account switching, and sign-out update only the user, account list, balance, usage, and model catalog. The current Workspace, Session, messages, composer content, navigation, and scroll position remain in place. Responses from a previous identity cannot overwrite current remote data. Saved Session model choices are preserved; if the selected model is absent from the new Federation, choose an available model before sending.

Account activity

The account page is centered with a maximum width of 768px and adapts to the window width.

After sign-in, the Settings sidebar account entry shows the current user's name and a 32px avatar in a 56px row, falling back to email or user ID if no name is available. Missing or failed avatars show the first letter of the name. When signed out, the entry shows the user icon and Account label. In Account details, click anywhere on the Federation address, user ID, or current account name row, or activate it with the keyboard, to copy its complete value. After copying, a brief success message appears in place of the value; failures show a message. Empty fields cannot be copied.

Select the account entry in the Settings sidebar (your avatar and name after sign-in) to update your balance, usage totals, annual heatmap, and trend chart together. Selecting the active Account item also refreshes the data. The first load shows structural placeholders for balances, usage totals, and charts. During a request, a small loading spinner appears to the right of the account name in the sidebar, keeping the avatar and name visible. After 200ms, a low-contrast sheen passes over the data cards. Fast requests do not trigger the card effect, and duplicate requests are ignored. Refreshes keep existing data fully readable; updated values briefly fade in while the selected period, metric, and chart position stay unchanged. With the system's reduced-motion preference enabled, the sidebar loading icon stays still and a static thin line on the data cards indicates loading. If balance or usage retrieval fails, its existing data is retained and an error is shown; select Account again to retry. Successful parts update normally. Requests from a previous account cannot overwrite the current account after login or switching.

The single Spending / Tokens switch at the top of Personal usage controls both the activity heatmap and trend chart; spending in USD is shown by default. The trend follows the Day / Week / Month selection, covering the last 30 days, 12 calendar weeks, or 12 calendar months. Missing dates count as zero. The Tokens trend shows total tokens with compact axis labels; hover, focus, or tap a data point to see its date range and exact count.

In Settings → Account, the Personal usage → Activity heatmap shows the past year. Spending mode colors cells by daily spending; Tokens mode uses daily total tokens. Zero values stay at the lowest intensity. Each day keeps the same cell size and spacing; scroll horizontally in a narrow window. Month labels stay within the chart edges. The heatmap initially scrolls to today, which has a border marker. Switching metrics, refreshing data, and resizing the window preserve your browsing position.

Hover over a day, focus it with the keyboard, or tap it to open a compact Downcity-style tooltip with contrasting colors, small rounded corners, and an arrow pointing to the date cell. The date and weekday appear at the top. A larger value and icon highlight the selected metric, with the other metrics and their icons arranged compactly below. Spending uses USD; token and AI call counts show exact values. The tooltip sizes to its content and uses a brief fade, disabled when reduced motion is preferred. Days without activity show only the date and No activity; days with calls and zero spending still show their call count. Dates follow the calendar days in the usage summary. Details stay within the window, and hovering highlights the cell border. Press Escape or tap outside to close details. Use Tab to enter the heatmap at today, or the last date if today is outside the range. Opening the page preserves the existing keyboard focus. Use arrow keys to move by day vertically or by week horizontally, and Home / End to reach the first or last date.

Create an Agent

On the Create Agent page, describe the work in ChatInput and choose the generation model in its toolbar. Click the round send button at the bottom right to draft a name, description, and instructions. Enter sends a single plain-text paragraph; Shift + Enter adds a new line. Use ⌘ / Ctrl + Enter to submit multiple paragraphs or code. Configure manually opens the draft directly. This entry supports text without attachments, references, or chat commands.

Edit the name, optional description, instructions, and default runtime model. Both model fields use the shared chat model menu with name search, selection marks, and model detail previews. Agent creation uses the model's default reasoning configuration. Generation and runtime model choices are kept separately; refreshing the catalog does not replace a selected model. Submissions pause while models are loading or when the selected model is unavailable. If no text models are available, configure one in model settings first.

Edit request at the bottom preserves the request and draft. Generating again asks for confirmation before replacing text you have edited; Continue editing draft returns to the existing content. Create Agent registers the Agent only after confirmation through that button. Failed requests show the reason and retain your input for a manual retry. The full request stays in the editor while generation runs, with editing, model selection, and sending disabled. It remains available after failure or when returning to edit. Duplicate actions are disabled during requests; leaving the page does not cancel a generation or creation request already submitted.

Agent profile

Right-click an Agent or Group row in the Agents sidebar, or click its ⋯ button, to open its settings or delete it. Deletion requires confirmation. The Agent or Group row stays highlighted while its context menu is open and returns to its previous state when the menu closes. The current conversation stays selected.

Open configuration from an Agent's menu. The BayBar tab uses segmented navigation for Identity, Model, SOUL.md, and sections such as Skills and Memory contributed by enabled Powers.

Configuration, files, attachments, references, composers, and Power resource details share one BayBar content container. Multiple sections show segmented navigation; a single section displays its content directly. Reopening a tab preserves its content and selected section, and an entry can request a specific section. A requested Power section stays selected while it loads; manual changes during loading use the latest choice. Removing a Power or invalidating its version immediately releases its content. Loading errors appear locally while core Agent settings remain available. A deleted Agent shows an unavailable state. Hidden tabs and sections stay outside focus navigation and cannot provide header actions; closing a tab releases its content. The composer hands the draft between the main view and BayBar according to the panel's actual visibility.

Power pages have transparent outer backgrounds and inherit the BayBar surface. Skills, MCP, and memory lists retain their group backgrounds. Text action buttons such as Refresh, New, Add, and Remove use a 32px height for easier clicking. The memory list header contains Refresh and New actions, with the count beside the list title. Skills starts with its search field.

Click a Skill or memory entry to read it in an independent BayBar tab. Opening the same resource again activates its existing tab. Closing a detail returns to its source tab when available, otherwise to an adjacent tab. Visited Power lists in Profile retain search and scroll position across section switches, tab switches, and panel collapse; unvisited sections load on demand. Closing the Profile does not close its details; deleted resources and unavailable Powers show their current state. Tabs remain within the current Desktop window and are not restored after restart.

Input, Textarea, and editor placeholders share a lighter color than entered text, adapting to light and dark themes.

Runtime input

Runtime inputs such as completed commands, task results, and tool attachments appear in arrival order inside Agent messages. Each input uses the same compact, transparent collapsed row as tool and reasoning items, with matching typography, spacing, and expand arrow. The row shows an icon, a readable title such as “Command result”, “Task result”, or “Attachments received”, and secondary counts of nonempty text/context blocks and files.

Expand a row to read the original text and view attachments. Text preserves line breaks and wraps naturally; command output uses a monospace font. Expanded content uses the standard indented tool detail container, with one scrollable area limited to 240px. Context labels use secondary text. Empty inputs display “No displayable content.” Open Source beneath the expanded content to inspect the event type and any available Tool Call ID or Turn ID. Attachments load only after expansion. Copying or sharing Agent reply text excludes runtime inputs.

Continuous Agent replies

Retry appends to the original canonical Agent Message in the same Turn. Completed text and tools stay in place. Copy, quote, share, and export use that message; new user input establishes its normal reply boundary. Background results do not appear as user messages, and forks retain their actual message anchor.

Completed messages without displayable content do not leave an empty avatar row. Running replies still show working status when reasoning is hidden. The activity line distinguishes model thinking, the tool being executed, and waiting for input. Choosing Stop shows “Stopping” until execution finishes its cleanup. A failed stop displays its reason and restores the controls while keeping the draft. With text in the composer, Stop remains available alongside Send. You can keep editing while stopping; sending and repeated stop requests are temporarily disabled.

Model execution, including internal automatic attempts, uses normal thinking, output, and Stop controls. A final request failure shows the current error and Retry.

Avatars

The Agent catalog reads identity summaries on demand. Opening configuration loads the complete instructions. Avatars load through controlled local file URLs; editing or replacing an avatar automatically selects its new file version.

Desktop uses circular Agent, group member, account, and placeholder avatars. Chats, sidebars, tasks, notifications, configuration, avatar selectors, and BayBar share the same avatar shape while keeping sizes appropriate to each layout. Groups retain their member composition, with each member shown as a circle. Uploaded images are cropped to a circle for display.

System requirements

The macOS app requires macOS 13 Ventura or later.

Local Workspace file discovery (find) and content search (grep) require ripgrep (rg) on the Desktop process's PATH. A missing executable returns rg_unavailable. On macOS, install it with brew install ripgrep.

Empty states

Sidebars, BayBar, task views, file previews, model lists, and Power pages share a centered empty state: a line icon above a short status label, with a description and action when needed. Empty panels center the placeholder within the available space; list sections and popup menus use compact spacing.

An expanded Workspace without conversations shows a conversation icon and “No conversations.” When all its conversations are pinned, it shows a pin icon and “All conversations are pinned.” Searches without matches retain their search-specific message. Loading and read failures show their respective status.

Workspace descriptions

The Workspace list reads registration metadata only. Desktop loads README.md when you open project details or the README editor; refresh retrieves the content again. Complete README content is limited to 2 MiB. Larger files show an error and remain unavailable in this editor; use the project file editor to work with the original file.

Loading local media

Local resources in chat and Workspace previews stream on demand and support a single byte range per request. Audio and video seeking can read only the requested interval. Cancelling a request or quitting Desktop closes its file handle. When multiple avatars or media resources load together, requests wait in arrival order for an available read slot. Cancelling a request or quitting Desktop removes its pending entry. Requests receive a busy response if the waiting queue is full or the wait exceeds 15 seconds, and can be retried later. Local resources do not use the 20 MiB limit for buffered remote resources.

Remote SSH resources still load one complete file of up to 20 MiB using the existing remote connection and error handling. These rules do not change the separate limits for Workspace file tools or the README editor.

Workspace file discovery

For find, path selects the search directory and pattern is a POSIX glob relative to that directory. *.ts matches only the current level; **/*.ts includes subdirectories. Brace patterns and extglobs are supported. Results include dotfiles, respect applicable ignore files (including .gitignore, .ignore, and .rgignore), exclude .git contents, and do not follow symbolic links. Selecting a subdirectory does not bypass project ignore rules.

File enumeration runs in a separate process and stops when cancelled. max_results bounds the returned list; an additional match sets truncated: true. Returned paths are relative to the Workspace root.

Background results in conversations

Work results appear in place as completed, failed, or stopped. Expand a result when it contains a deliverable; command results show actual output and necessary failure reasons. Routine internal maintenance and duplicate media intended for the model do not occupy message rows. Results omit machine envelopes, source identifiers, and content counts.

Each Power describes its own business phase: work submitted, memory processing or saved, task execution queued, a message queued or delivered, or image generation in progress. Only useful bodies expand. Session uses the same conversation bubble icon as Chat, and switching language updates activity wording.

Inbox and unread indicators

Inbox uses Downcity's borderless page layout and subtle selection surfaces. Agent avatars appear at the start of list items and beside source names in details, using Desktop's current Agent directory. Missing avatars use the default Agent icon; Power updates without an Agent use a business icon. Question, approval, and private input forms work directly in details, with matching titles shown once. Refresh preserves input; switching or closing asks before discarding unsubmitted edits.

In chat and Inbox, enter private answers in masked fields and click Submit response. Submitting a value with the same name updates it in the current Session, and later uses of that reference receive the new value. Completed forms hide private answers; the Agent receives only a reference.

The Inbox Sidebar offers Needs attention (respondable requests and unread failures), Notifications (ordinary notifications), and All (all Inbox items, deduplicated). The desktop list stays 360px wide at the same position before selection, while an item is open, and after closing it. Its title, count, refresh, organization menu, and messages share one subtle panel. Each item has a separate 32px avatar column. The first text row shows the source name followed immediately by the time, the second shows the actual notification title, and up to two lines summarize the real body below, with empty summaries omitted. The attention dot stays at the upper right. Notification menus use a separate lower-right slot and support hover, keyboard focus, and touch without covering text or the dot. Both toolbars are 36px high. The right area stays blank until selection; detail bodies and forms fill the remaining width with consistent padding, and form actions use shared 32px primary and secondary buttons. Select an item to open details on the right in a wide area, or replace the list with details and a Back action in a narrow area. Entering Inbox does not automatically select an item, and clearing the selected update does not open or mark the next item as read.

Desktop includes one City-level Notification Power. Inbox, conversation indicators, navigation dots, and the application icon badge share its state and refresh when data changes. The badge and Inbox navigation number count respondable pending requests plus unread Inbox failure matters. The sidebar and navigation tooltip show both components. Success and informational updates retain unread dots without increasing this number, while the Notifications category separately shows the unread update count. Ordinary Agent replies only create conversation unread indicators; background Work Main results, failures, and original requests needing a response enter Inbox. Each conversation, group conversation, or Task updates one matter, while execution history stays with its source.

Viewing a question or approval does not finish it. A focused, visible source conversation marks ordinary updates as read; Inbox offers notification row menus to mark or clear ordinary updates, plus list header menu actions to mark all updates as read or clear read updates. Both header actions apply to ordinary Inbox updates across all categories in the snapshot at the time of the click; clearing read updates only affects versions that remain read. Actions match IDs and content versions, preserving new matters and newer versions. Marking read keeps records; clearing only removes ordinary notifications. Later events for the same matter can notify again. Source history and original questions, approvals, and private input requests remain intact. Private answers use the original Session's dedicated submission channel. Archived or removed sources retain notification summaries and read state, with an explicit explanation when they cannot be opened. Reading notifications never deletes history. See Notification and Inbox.

Tasks in BayBar

The Task Sidebar offers Memo, Needs attention, and Tasks. The main area browses List, Board, Calendar, and Timeline. Each task opens one stable BayBar tab with Task, Delivery, and History as applicable. New task and Edit task retain drafts across tab switches and chat navigation. Natural-language organization fills a suggestion for review before saving. Immediate execution requires Save and run. Closing a dirty tab offers continued editing or explicit discard.

Task system notifications

While Desktop is running, personal reminders, successful Agent or script executions, results awaiting acceptance, execution failures, triggers that cannot start, and newly discovered missed one-time arrangements request an operating-system notification. Each execution sends at most one final result notification: “结果待验收” for results requiring human acceptance and “执行成功” for other successful executions. Creating, editing, starting, pausing, resuming, and skipping work do not send system notifications.

A notification activates its task's BayBar tab and selects the specified execution, including when that tab is already open. The main collection stays in place. Desktop retains the latest click while hidden, minimized, or preparing. Deleted sources show an unavailable message; failed reads offer retry. Opening a delivery only views it. Acceptance, feedback, ignoring a failure, and marking an update read remain explicit actions.

Allow notifications for Downcity in your operating-system settings. Permissions, Do Not Disturb, and system policy can suppress banners; tasks and in-app pending matters remain available. Automatic arrangements require Desktop's City to stay running; a fully exited app cannot wake the system. One-time arrangements are checked by the minute and become missed matters after the existing tolerance. Delivery is best effort: restarting does not replay historical notifications, and notifications may be lost during a process crash.

Session models

The new-conversation page only asks you to choose an Agent. Model and secure mode controls, including their slash commands, are available after the Session is created. Creation uses the selected Agent's model and reasoning settings. When that model is unset or unavailable, Desktop uses the default text model in Settings → General, then the first available text model in the catalog. The initial secure mode requires approval before execution.

When you reopen a conversation, Desktop uses the model saved for that conversation. If the menu says Select a model, choose one before sending. Your choice is saved for the next time you open the conversation. Desktop checks the model configuration before sending; if a model is missing, it keeps your message and prompts you to select one. Queued messages are also kept, and automatic sending pauses after a failed attempt. Select a model, then resume sending.

Start a session from a Workspace

In the Works sidebar, choose New conversation on a Workspace row. The draft inherits that Workspace. The page header shows the current directory; hover to view its path, or click to search by name or path and switch Workspaces. Switching before the first message preserves the complete text, attachments, and references.

When the draft has no Agent pill, the composer automatically opens the Agent candidates. Choose an Agent before sending. The owner appears as an inline pill matching references and attachments, with the Agent's avatar, name, and remove button. Agents without a custom avatar use the default Agent icon. Type @ to replace the owner. Removing the pill opens the Agent candidates again and preserves text, attachments, and references. When you type @ manually, the menu also offers files from the current Workspace; choosing a file does not change the owner. Use arrow keys to move through candidates and Enter to choose. Confirming text with an input method does not submit a message.

Sending the first message creates a Session owned by the selected Agent. Switching a draft's Workspace or owner closes its expanded input panel and returns the complete draft to the current page, where you can continue writing and expand it again. Created Sessions retain their owner. Start a new conversation from a Workspace to work with another Agent.

Inputs and focus

Creation pages, settings, and task forms use standard field styling with compact, standard, or large sizes. Search icons and text form one field with shared height, corners, padding, and focus feedback. README, environment configuration, and group instructions use document layouts suited to longer writing. Chat keeps its card corners, historical message editing keeps its surface corners, and the model toolbar button keeps its pill shape.

Fields use a background one shade darker than their container across pages, sidebars, BayBar, cards, and dialogs, without borders or inset outlines. Hover and focus progressively deepen the background without an outer ring extending beyond scrolling edges. Search groups draw one field background, while rich-text editors retain their own card shape. Chat composers and historical user-message editors both have a subtle background at rest, so the editing area remains visible without hover or focus. Read-only content remains selectable and copyable, disabled fields appear subdued, and validation messages appear beside the field.

Task and settings selectors share menu styling and support keyboard navigation, selection, and dismissal. Radio options, checkboxes, and note task lists use consistent selection marks with visible keyboard focus and disabled states.

Pill switchers in Memory, Agent configuration, and settings size to their option content. On first display or when reopening a panel, the selection background appears directly under the current option. It slides when you switch options and aligns immediately when the layout or size changes.

Chat drafts are saved automatically on this device after you pause typing. Leaving the input area or switching sessions immediately saves any pending changes. Each session restores its own draft when you return. Expanding or collapsing the right-side input panel preserves the current draft during the editor handoff. The main conversation input takes over as soon as the panel starts closing.

Text selections use a subtle shared highlight across inputs, documents, messages, and file previews, preserving the original text color with the same highlight strength in light and dark themes.

Local diagnostics for UI stalls

Open Settings → General → Performance diagnostics and turn on Enable diagnostic logs to continuously record UI stalls and main-process performance. The switch is off by default, applies immediately, and resumes collection on the next launch when enabled. Use Open log folder to view daily YYYY-MM-DD.jsonl files. The default directory is ~/.downcity/logs/desktop-diagnostics/; with a custom local data root, logs are in its logs/desktop-diagnostics/ directory.

Logs retain today and the preceding 6 local calendar days, with cleanup at startup and when the date changes. Turning collection off leaves existing logs subject to retention. Cleanup cannot run while Desktop is closed and catches up at the next launch. Each daily file is limited to 20 MiB. Once full, it continues counting dropped events; event recording resumes the next day. The settings page displays write failures, cleanup failures, and size-limit status.

High-frequency events are aggregated every 30 seconds; events taking at least 50ms retain details. Stalls lasting at least 500ms are submitted promptly after recovery, and a full detail batch is also submitted early. Resource records retain the window's latest main-process sample and CPU and memory peaks. Writes are asynchronous and batched; a busy event loop or disk can delay them. Both sending and writing queues are bounded. Dropped counts appear in batch dropped_events or the final logger.dropped statistics line. Turning collection off may discard unsubmitted renderer windows or queued batches while counting the loss. Normal shutdown submits the final main-process window. Force-quitting, power loss, or a renderer crash can still lose data that has not been submitted or written.

Continuous logs do not contain message content, actual keys, or source paths, are never uploaded automatically, and do not enable CPU profiling. They are independent of the manual in-memory report below: stop() does not turn off logging enabled in settings.

Manual diagnostic reports

To export detailed records from a short recording, run the following command in Desktop developer tools Console. Each start() clears the previous manual report:

await window.downcity_diagnostics.start();

Return to the interface and reproduce the stall while typing, switching sessions, scrolling, or selecting text. Once the interface recovers, promptly return to the Console to stop recording and download downcity-diagnostics.json, so later events do not overwrite the period you need to inspect:

await window.downcity_diagnostics.stop();
await window.downcity_diagnostics.download();

Use await window.downcity_diagnostics.snapshot() to inspect the combined report. The renderer and main process each record for up to 3 minutes and retain their latest 2,000 events plus aggregate counts for the entire recording window. A busy main process can delay stop requests and automatic timers. Manual reports stay in memory, cannot be retrieved after refresh or restart, and are never uploaded automatically. Diagnostic data contains only fixed operation names, numbers, booleans, durations, and anonymous identifiers scoped to this recording, without input text, actual keys pressed, paths, or original session IDs.

Renderer recording covers long tasks, input queueing, time from input to the next requestAnimationFrame callback, visible-page frame gaps, editor transactions and restoration, streamed message batches, message and segment projections, session snapshots, segment reclamation, and width and scroll changes. The report's main section contains main-process event-loop delays, CPU usage, process memory, garbage-collection durations, and streamed message send durations. If main-process collection is unavailable, main is null and capabilities.main is false. CPU and memory metrics cover the whole main process and cannot be directly attributed to one Agent.

Session-switch records also include synchronous navigation notifications, listeners taking at least 8ms, navigation persistence, message-cache trimming, and synchronous editor Hook work. Where supported, the report includes long animation frames, forced style/layout time inside scripts, and React render durations for the sidebar, main view, messages, and composer. capabilities.react_profiler becomes true only after an actual React callback during recording. Standard production React does not provide this callback; false does not mean rendering has no cost. React render durations exclude DOM commits, layout effects, and browser painting, and nested boundaries must not be added together. The interval from style/layout start to the end of a long animation frame includes subsequent work; it is not pure layout or paint time.

Each process records at_ms relative to its own capture start. Use started_at_epoch_ms + at_ms for approximate alignment. Event-loop delay is estimated with a 250ms heartbeat and recorded only when it reaches 50ms; system sleep and scheduling can also increase it. This metric alone cannot identify the code responsible for a stall. Garbage-collection monitoring depends on runtime support, which is marked in capabilities. Time to the next frame callback is not an exact measurement of actual painting, FPS, or INP. Diagnostic durations can overlap through nesting and must not be added together.

Deep stall diagnostics

Enable Deep stall diagnostics under Settings → General → Performance diagnostics to automatically retain main-process and renderer CPU profiles around significant stalls, together with backend stage timings. This separate switch is off by default and resumes on the next launch when enabled. Raw profiles contain function names and source paths, stay local, and are never uploaded automatically. Continuous sampling adds CPU, memory, and debugging-connection overhead; turn it off after investigating.

The requested sampling interval is 5ms, with windows rotating approximately every 30 seconds. When a main or visible-renderer heartbeat is delayed by about 1 second, diagnostics attempt to save the previous completed window and current window. Triggers have a 30-second cooldown and limits on targets, memory, and artifact size. An independent Worker monitors the main heartbeat and attempts to persist a stall-start record even while main is blocked. Recovery adds a recovery record and saves available CPU profiles. Renderer detection runs only while the page is visible and reports after recovery; a blocked main process can delay delivery, while the original measured interval is retained.

Settings display actual main and renderer sampling status, independent-monitor status, saved incidents, and failures. Developer tools or another debugger may make renderer sampling unavailable; existing debugging connections are not taken over. Manual CPU recording pauses main rolling sampling and resumes it afterward. Rotation and pauses create gaps recorded in incident metadata; a gap does not mean no stall occurred.

Artifacts are in deep/YYYY-MM-DD/ under the diagnostic log directory and retain today plus the preceding 6 days. The combined daily deep-artifact budget is 100 MiB, including 1 MiB reserved for independent heartbeat records; each CPU profile is limited to 15 MiB. Drops due to limits, failures, or cooldown are counted. Incident JSON links neighboring .cpuprofile files and backend stages. Load CPU files in developer tools. Selected IPC calls record request IDs, approximate dispatch-to-receive waiting, and host-handler elapsed time. Agent stages use anonymous IDs for tool, model-request, and execution boundaries. Async elapsed time includes waits and scheduling, is not CPU time, and does not imply an inferred causal parent between an IPC call and a background Agent.

CPU profiles identify JavaScript/V8 execution hotspots but cannot guarantee native-wait, GPU, or kernel stacks. Force-quitting, unavailable sampling targets, or startup timing can leave only an independent heartbeat record without a CPU file. Interpret status, gaps, and drop counts together; missing samples do not prove normal operation.

Main-process CPU profiles

For application-wide stalls lasting several seconds, enable CPU sampling separately to identify the functions executing in the main process. Raw profiles contain function names and source paths. They stay in local memory and are never uploaded automatically; inspect local paths before sharing. Sampling uses a local inspector Session without opening a network debugging port.

await window.downcity_diagnostics.start_cpu();

Reproduce the stall, then stop and download downcity-main-ID.cpuprofile once the application recovers. download_cpu() stops sampling first, so it can also be called on its own:

await window.downcity_diagnostics.stop_cpu();
await window.downcity_diagnostics.download_cpu();

CPU sampling is independent of regular diagnostics. Its target interval is 1ms, and it stops automatically after about 3 minutes while retaining the result. A blocked main process can delay stopping. Sampling adds overhead and is intended for short reproductions. Reloading, restarting, closing the page, or a renderer crash releases sampling and its in-memory result. Starting a new capture replaces the previous result. Import the file with Performance → Load profile in developer tools. To compare sidebar timings at the same time, also enable regular diagnostics and download both reports separately.

Test System One decision models

Open Settings → Models, select JEV or another model in the decision model section, and use Test decision model in its details. A ready-to-run example is prefilled. Edit the content and decision instructions, choose Boolean, Choice, or Score, then select Run test. Choice accepts one distinct option per line (2–255 options); Score requires a minimum and maximum with a range width of 1–9. Update the instructions to match the selected answer type.

The result shows structured answers, any returned probability or option distribution, usage when provided, and elapsed time. Failures appear in the same form so you can correct the input and retry. Tests use the current account and the model's normal billing. Closing the details does not cancel a request already submitted.

Workspace home

Workspace home shows a project icon beside the Workspace name, its path below, and the rendered README as the page content. SSH projects also show the remote host. Hover over the title row or focus its controls with the keyboard to reveal the edit button on the right. The more menu provides editing and copying the path, with system file manager access for local projects and refresh for SSH projects. Editing opens the Workspace configuration in the right panel.

Add a local or SSH project

Use the + menu in the upper-right corner of the Works sidebar and choose Add local folder or Connect remote SSH. The command palette and the empty-state add button open the same local folder dialog. Click the Choose folder area in the dialog to open the system picker. Once you select a folder, the dialog shows a folder row and an editable Name filled in with the folder name. Click the folder row to choose a different location. Cancel and Add appear at the bottom; click Add to register the folder and open its Workspace home.

Cancelling the system picker keeps the dialog open. Closing the dialog before clicking Add leaves workspaces unchanged. If adding fails, the dialog shows the reason and keeps your folder and name so you can try again. Adding an already registered folder opens its existing Workspace and preserves its name.

Choose a concrete SSH host alias from your local ~/.ssh/config. The folder picker opens the remote login user's Home. Click a folder to enter it, or use the up arrow to go to its parent. Use New folder beside the path to enter a name inline. Press Enter or click the checkmark to create and enter the folder; Escape or the cancel icon discards the name. Creation errors keep the input available for correction.

You can also choose Enter path manually and type an absolute path. Submitting checks the path first. If it does not exist, the same form offers Create folder and continue, which creates missing parent folders, then proceeds with the selected dependency installation and registration. Checking or creating a manual path does not require Python. Editing the path clears the creation confirmation. Created directories remain if a later step fails or you cancel registration. The directory name becomes the Workspace name.

Folder links have a link icon; entering one shows its real location, which is also used for registration. The root has no parent. Lists show at most 100 folders per page within a 64 KiB response budget; use Next folders to continue and refresh to return to the first page. Each request scans at most 20,000 entries. Permission, connection and output-limit failures show a reason and Retry. Changing hosts clears the selection; navigating or closing the picker cancels previous browsing requests. While entering a new folder name, finish or cancel that input before submitting the Workspace. Creation locks navigation and host changes. Registration disables repeated submission. Interrupted directory creation is not automatically replayed; inspect or refresh the remote directory before retrying.

Read-only browsing uses system OpenSSH with a temporary static snapshot of user and system configuration, retaining user, port, keys and ProxyJump settings. It disables LocalCommand, RemoteCommand and KnownHostsCommand. Configurations containing Match exec, custom ProxyCommand or dynamic Include paths are rejected before connecting; use static Include paths and ProxyJump. Browsing requires Python 3.8+ on the remote host. If it is missing, enter the path manually.

The dropdown reads configuration files without running SSH or configuration commands. It supports multiple aliases per Host line, quotes, comments, optional =, and nested unconditional Include files with ordinary *, ? and bracket globs in lexical order. Relative Include paths resolve under ~/.ssh. Exact duplicate aliases are removed; spelling and case are preserved for OpenSSH; wildcard and negated Host patterns are omitted, as are concrete aliases excluded by a negation on the same line. Selectable aliases use letters, digits, _, ., : and -, starting with a letter, digit or _.

Include directives inside conditional Host/Match blocks and paths using environment variables, % tokens, ~user, braces, POSIX named character classes or remaining backslash escapes are skipped with a notice. Host * and Match all count as unconditional. Put selectable aliases directly in ~/.ssh/config or an unconditional Include. System-wide configuration still applies to connections, but is not enumerated. Configuration files must be UTF-8. Missing configuration, no selectable aliases and read failures have distinct messages and a Reload hosts action. Cycles, invalid relevant syntax, non-regular files or file/size/depth/directory limits fail without returning a partial list.

SSH Workspaces run on Linux or macOS and use Python 3.8+, ripgrep (rg) and POSIX ps. Select Install missing dependencies when adding the Workspace to install and verify these tools on the remote host before registration. For an existing SSH Workspace, choose Install missing dependencies from its actions menu. Downcity checks the host identity and directory first, skips tools already available, and reports installation failures. It does not automatically retry an interrupted installation.

Installation uses an existing apt-get, dnf, or apk on Linux, or Homebrew on macOS. Linux installation needs root or non-interactive sudo. If the package manager or privileges are unavailable, have an administrator provide the missing tools on the remote SSH PATH. Desktop uses system OpenSSH, system keys and the existing SSH agent. It does not forward the SSH agent or start a persistent remote service.

OpenSSH strictly checks known_hosts. For an unknown host, Desktop displays OpenSSH's address, key algorithm and SHA256 fingerprint. Verify the fingerprint with the host owner and explicitly trust it to continue. Changed host identities are rejected. The SSH agent is not forwarded to the remote host.

Files, search and commands operate on the remote machine, without a local project checkout. Agents, Session history and attachments remain with the current Desktop. File tools are rooted at the registered directory and reject symbolic-link escapes. Shell commands run with the remote login user's permissions, without local Sandbox isolation; commands can access resources available to that remote user.

Folder and connection icons identify the source in the sidebar; details show the host and remote directory. README, text, image and attachment access use controlled file capabilities. Text previews and explicit media reads are each limited to 2 MiB. Use Refresh in remote file menus; remote files do not have local live watching. System applications, Finder and local VS Code actions are offered for local projects. Remote file links carry the Workspace ID so identical paths on different hosts stay distinct.

A failed connection preserves the registration and its name without hiding other projects. Refresh to retry reads. A disconnection after submitting a command or modification can leave its remote outcome unknown; Downcity never automatically replays it. Inspect remote state before retrying. Removing a registration preserves remote files and local history. Edit names through Workspace configuration; local SSH identity authorization is presented separately from remote project connections.

Local City management

Desktop owns one data root under an exclusive lease acquired before database initialization and Agent assembly. While it is the current host, city agent, city workspace and city session connect to its verified local management channel, independently of the City Remote HTTP switch. Do not start a second host for the same root. Agent defaults affect new Sessions; existing snapshots and model overrides remain. Workspace removal preserves project files and history and refuses active execution or Shell processes. See Resource Management.

City Remote

Open Settings → City Remote to enable HTTP access, initially bound to 127.0.0.1:5314. The switch applies immediately; save changes after editing the address or port. The page shows the actual running status, connection URL or startup failure. Use a connection code in Remote or Chrome Extension; there is no need to run city on separately.

Optionally save a Remote HTTPS address on this page after configuring your own Cloudflare Tunnel or reverse proxy. Saving or clearing this address only changes the default for future connection codes; it does not restart the listener, interrupt connections, revoke credentials or rewrite previously shared codes. The address is checked for format only, without a network request or proof that a Tunnel reaches this City. Desktop does not run cloudflared or manage Cloudflare credentials.

Select Create connection code on the same page. Its address defaults to the saved HTTPS address, or the actual local listener when the setting is empty; you can override the address for this code. Enter a recipient note and a City address reachable by that recipient, choose the allowed Agents (or explicitly all existing and future Agents), and set an expiry of 30 days, 90 days, no expiry or a custom date. Each creation issues an independent token. Copy the complete connection code from the one-time display and send it privately. Closing the dialog clears the code. Existing authorizations can be edited or revoked, but their plaintext cannot be recovered. CLI city token manages the same authorization records.

In Remote, paste the code into the single connection-code field. Remote decodes it locally and shows the City address for confirmation; check it, then click Connect. Remote connects directly to City without a Downcity backend relay. It does not read codes from the page URL or save codes or credentials; only a non-sensitive City address may be remembered. Reloading requires the code again.

The connection code contains an address and token, encoded rather than encrypted. Anyone holding it can use the authorized Agents until expiry or revocation. The recipient note does not bind access to a person's identity or prevent forwarding. Revoke the authorization if the code is exposed; create a new code if its plaintext is no longer available.

To use Remote on the same computer, create the code with http://127.0.0.1:5314 (or the running address shown in settings), then open https://remote.downcity.ai and permit local network access if the browser asks. The site permits HTTP only for exact loopback hosts 127.0.0.1, localhost or [::1]; the target must match the actual listener. A local development Remote at http://127.0.0.1:5173 or http://localhost:5173 is also supported. For another computer, use an address reachable from the recipient's browser. Desktop allows both local Remote origins and https://remote.downcity.ai, so these origins need no additional CORS configuration. For a City on another device, configure a reachable HTTPS endpoint with a trusted certificate yourself. The browser-origin allowlist does not provide HTTPS, a tunnel, or public network access. See Remote for browser access requirements.

HTTP is disabled by default. Enabled settings are restored on Desktop startup. Disabling access, changing listen settings or exiting Desktop disconnects clients. Desktop and CLI retain the single-host model. Loopback is local-only; when binding another address, clients must use the corresponding machine address and network configuration.

Add files and paste images

Pasting a standalone local file:// address, such as file:///Users/me/project/reports/assessment.md, creates a file reference showing its icon and filename. Hover to see the full path. Clicking follows file navigation: registered workspace files open in a BayBar file tab, and other local paths open with the system default app. Percent-encode spaces and special characters in filenames. Missing or currently inaccessible files show a warning icon with an explanation on hover. The reference stores only the file address, which the Agent can read within its own permissions. Adding a reference or opening its preview does not deliver file contents to the model. Dragging or selecting files adds session attachments. File addresses mixed into prose stay as text.

In ChatInput, press ⌘ + Shift + V (Ctrl + Shift + V on Windows / Linux) to paste without formatting. Only clipboard plain text is inserted, preserving line breaks and blank lines. Code fences and Markdown markers stay literal, references become readable text, and images or files are not added as attachments. Pasting inside a code block preserves plain text and line breaks.

Attachments in the composer, pending queue, and user messages share a plain inline capsule with only a small icon and filename. Its font matches references at 87.5% of the surrounding body size, with regular weight and an upright style, and its height matches the text line height. Attachments and references follow the body baseline: filenames align with adjacent text, while icons are vertically centered. Long filenames truncate; hover to see the full name. Clicking an attachment opens its own BayBar preview tab. Reopening the same file selects its existing tab, and switching conversations keeps the preview available. Images offer fit-to-width and original size; audio and video use playback controls. Markdown, static HTML, CSV/TSV, and text support previews. Text previews are limited to 2 MB. Unsupported files and read failures show a message; saved local files can be opened through Open with default app above the preview. Draft attachments can be previewed before sending, with a separate remove button.

Content in the composer, pending queue, and user messages shares the conversation font family, line height, and text color, and responds to the conversation text-size setting. User-message bodies and collapsed previews use compact sizes: 13px for Small and Standard, 14px for Large, and 15px for Extra large at the default interface scale. Pending-message summaries use 87.5% of the conversation body size, matching the text in the queue's attachment and reference capsules. References use the same rounded capsules across all three surfaces, with a neutral background that adapts to light and dark themes. Each capsule shows a small icon and a single-line excerpt at 87.5% of its surrounding body text size, using regular weight and an upright style while inheriting the shared body typography. The excerpt aligns with adjacent text on the same baseline, while the icon is vertically centered within the capsule. Capsules are limited to 24rem wide, shrink with their container, and truncate long excerpts with an ellipsis. Click to read the full reference in its own BayBar tab. Reopening the same reference selects its existing tab; different text from the same source opens separately. Tabs keep the full content captured at the time of opening, even after switching conversations or changing the source. The composer has a separate remove button; the queue and user messages share the same preview action. The full reference is still sent with the message. Opening references or attachments does not submit the draft.

Files selected locally, pasted images, and workspace files are added to the input area that started the read. Switching sessions, expanding or collapsing the input panel, sending a message, using /clear, or restoring input content externally cancels the pending addition. Select the file again in the target input area when needed. Continuing to type in the same input area does not cancel the read.

The per-file upload limit is 20 MiB. Draft attachments are stored locally in IndexedDB as Data URLs, used only for drafts and transport. On submission, the Session saves attachments in its own attachments directory, and canonical file Parts record their paths.

Attachments always provide paths and metadata. Static PNG, JPEG, and WebP confirmed from file headers can also be included automatically when each is at most 2 MiB, all request media total at most 4 MiB, and the model input budget allows it. Recent images take priority without changing message order; current and historical attachments follow the same rules. PDFs, other files, and images omitted for budget remain references for the Agent to read when needed. Opening a BayBar preview does not itself deliver content to the model. Text reads return at most 64 KiB at a time. Explicit image/PDF reads have a 2 MiB per-file limit; oversized files return media_too_large and explicitly report no delivery. read provides no image conversion, PDF pagination, OCR, or Office parsing. Remote URLs are not implicitly downloaded, and one missing attachment does not block the entire conversation.

Edit historical user messages

After editing a historical user message, if the “Continue from here” selector opens, click an option to send using that action. “Create branch” is selected by default: it preserves the complete original conversation and continues with the edited content in a new branch. “Replace later messages” removes later messages and continues in the current conversation.

The selected option uses a light theme-color background and a checkmark on the right. Arrow keys change the selection; press Enter to confirm. A separate thin ring indicates keyboard focus. When the conversation has pending messages, replacement is disabled and its explanation shows why. Closing the selector preserves the editing draft.

StatusPanel above the composer

StatusPanel above ChatInput brings together queued messages, pending interactions, and child Sessions in three category pills: Queue, Interaction, and Sessions. Each pill shows a category icon, its name, and a separate numeric tag. Numeric tags use the same rounded shape, background, and font weight as the reasoning tag in the model selector. Only categories with content appear; the entire panel is hidden when all are empty. One-to-one and group chats use the same container. Group chats show only Interaction and have no local Queue or Sessions category. Conversation drafts also have no Sessions category. The main chat composer and the composer expanded into BayBar share StatusPanel.

Sessions counts all loaded descendants of the current Agent Session, including multiple levels of delegation across Agents. While any child Session is running, the pill shows a spinning loading icon; it returns to the hierarchy icon when none are running. The spinner respects the system reduced-motion setting. Expand it to see Session titles in their parent-child hierarchy. The left icon carries the state: a spinning loader while running, the hierarchy icon when idle, and the existing icons for pending or unread states. Rows do not repeat the state or show an Agent name on the right. The main chat composer and the composer expanded into BayBar share the same entry point: clicking a child Session opens it in a dedicated BayBar conversation tab while the main page keeps the current conversation. Clicking the same child Session again activates its existing tab. The tab shows that Session's messages, live execution status, and pending interactions, and supports loading earlier messages. BayBar child Session tabs omit ChatInput, Queue, and Sessions, while retaining Interaction for questions and approvals. The fixed bottom row centers larger action buttons. Reference stays visible with neutral styling, while Stop uses red text on a subtle red background and appears only when the child Session is running. Clicking Stop opens a confirmation popover above the button; execution stops only after confirming. Cancel, an outside click, or Escape closes the popover. Reference inserts a Session reference into the originating main conversation input captured when the tab was opened, without sending it. If the main page is showing another conversation, Reference first navigates back to the originating conversation and waits for its input before inserting. Repeated clicks are blocked while inserting. An active queue edit must be finished first; archived source conversations cannot receive references. Stop affects only this child Session. It uses the live snapshot when available and falls back to the lightweight runtime or directory status while synchronization is pending. Missing Turn information is read from the current runtime before stopping. Stop is hidden when execution ends and disabled while a stop request is in progress. The child Session list uses only the already-loaded active Session directory and does not automatically load the full history. The category is hidden when there are no loaded child Sessions.

Category pills and content share one continuous shell with a single background and rounded outline, without a gap between them. The pill group fits the available width and wraps in narrow windows without horizontal scrolling. Hovering a category button highlights its background; the selected button uses a stronger hover background. This feedback applies whether the panel is collapsed or expanded. When you enter an unarchived one-to-one main conversation with pending interactions, the panel opens directly to Interaction, including requests reported through descendant Sessions. It also opens when pending interactions first become available after asynchronous loading. Automatic expansion happens once per visit; manually collapsing the panel or selecting a category preserves your choice for that visit. Re-entering the conversation allows pending interactions to open again. Without pending interactions, the shell starts collapsed around the pills. Expanding widens the shell and grows content upward while the pills stay at the bottom; the selected background slides within the pill group. Click a category to expand its content upward, then click the active category again to collapse it. Switching categories shows only the selected category's content and smoothly animates the panel to its measured height. Queue changes and interaction form resizing use the same height transition without scaling text or controls. Content appears above the pills and scrolls within the area, with the pills staying next to the composer. When the active category becomes empty, the panel collapses and removes that pill; other categories with content remain available. After the first automatic expansion or a manual category action, new items only update counts. Animations follow your system's reduced-motion setting.

Queue and Sessions use shared list rows with 8px corners, a 6px outer inset, 2px between rows, and a minimum height of 32px. Multiline content grows naturally. Rows are transparent at rest; hover and queue editing use a light full-row background, without separators. Interaction forms keep their own layout, with option rows using the same corners and spacing.

Send follow-ups and take over

You can switch sessions or input panels while a submission is pending. Completion of the original request does not clear input in another session or take focus from another control you are using. Inline task input clears only after a successful submission; edits made while waiting are preserved.

While the window is temporarily hidden, direct and group chats continue receiving subscribed message updates in batches. Returning to the window shows the current progress. Expanding or collapsing the input panel does not change the pending-message list.

When you send a Chat message, it is submitted immediately only while the session is idle, the queue is not paused, and the pending list is empty. While a Turn is still running, or while the pending list already holds messages, new input is appended to the end of that list. While the queue is not paused, messages are submitted one at a time as each Turn finishes, without overtaking what is already queued.

Click StatusPanel's Queue pill to view queued messages. Each message shows its status and content on the left, with Send as steering and a Message actions menu on the right. The menu contains Edit, Move to top, and Delete queued message. Move to top puts the message first among the items after any active submission, preserving the other messages' relative order and the queue's pause state; it is disabled when the message is already first. Queue, Interaction, and Sessions each have a shared header at the top of their content, with consistent height, typography, and spacing. The category name is on the left and category actions are on the right: pause/resume for Queue, or navigation arrows and a page count for multiple Interaction requests. Interaction also shows the current source beside its name; Sessions shows its category name. Headers stay outside scrolling content and hide with their category when the panel collapses or switches. The main chat and BayBar composer use the same queue controls.

Pause and resume apply to the entire conversation's queue. Resume queue clears that pause: an idle session sends queued messages automatically in order; a running session waits for the current Turn to finish before continuing. Queues restored after a restart are paused as a whole and require manual resume.

Choose Edit from a message's menu to load its complete content into ChatInput, preserving formatting, attachments, and references. Your ordinary chat draft is kept separately and restored when editing ends. ChatInput uses a distinct background for edit mode and keeps the cancel and confirm controls. The message being edited is highlighted within the list's inner rounded edges, with a small pencil badge on the left to identify it while revising its content. Status icons share a fixed-size slot, keeping message text aligned when the state changes. Confirming updates the original queued message while preserving its position and the queue's pause state; it does not directly send the message. Empty content cannot be confirmed. Canceling leaves the queued message unchanged.

While editing, an edit lock temporarily blocks sending from the conversation's queue without changing the pause state you set for the whole queue. You cannot edit another queued message, manually send queued messages, change their order, or delete them, and the pause/resume button in the top header is disabled. Confirming or canceling releases the edit lock; automatic sending then follows the queue's pause state and the session's execution state. The main chat and expanded composer panel share the same edit. Collapsing the queue, switching conversations, or moving between the main chat and composer panel preserves the unfinished edit; return to that conversation to continue.

Items in the list are local records until they are submitted: you can edit, delete, or move them to the top. Option/Alt + Command/Ctrl + Enter appends the current input to the end and pauses the entire conversation's queue; Command/Ctrl + Shift + Enter bypasses the whole list and submits the current input immediately, preserving the remaining queued messages and the queue's pause state. If automatic sending fails, the original message stays in the list and the entire queue pauses. After resolving the cause, resume the queue or use Send as steering to explicitly send a chosen item. Manually sending that item does not resume the rest of the queue.

Each Turn is owned by its initiating actor. The same user can send follow-ups. When another actor owns the work, a queued submission is rejected, the original message stays in the pending list, and the entire queue pauses. To take over, stop the current work, wait for it to finish stopping, then send your message. Stop targets the Turn shown in the interface; refresh and retry if it has changed. A rejected submission does not mark existing work as failed.

Previously registered callbacks still deliver the stopped work's original result, never the answer from your new Turn.

Retry a failed model step

When a model request still fails after bounded automatic attempts, that execution ends as failed and releases its resources. The conversation shows one current error banner and Retry, with no running indicator or running count. Authentication, permission, parameter, content rejection, and other model errors all permit manual Retry; Provider retryability controls automatic attempts only.

Retry continues the exact failed model Step in the same Turn and original Agent Message. It adds no user message, inbox, or recovery prompt. Uncommitted output is discarded and completed tools do not run again. New tools bind to the current signal, approval, and output. The original method call retains its failed result; Retry has its own completion result.

Automatic queued messages continue in order after failure. New input or changed configuration can invalidate the old Retry. A failed child delivers its failure normally; the parent does not wait for someone to click Retry. A later retry does not reopen an ended parent task. An in-memory failed request does not prevent archiving or resource release.

The error banner uses a compact secondary Retry button with hover, keyboard-focus, and disabled feedback. Once Retry is accepted, the current error clears and ordinary thinking and output continue in the original message. Another failure replaces one current error banner; success leaves only the actual result. Automatic attempts use ordinary running indicators, without a countdown, attempt list, receipt, or success badge. Stop captures the currently displayed message checkpoint so a late request cannot stop a later same-Turn call.

Only one submission of a binding is accepted. If reception is uncertain, the app reads current execution state rather than sending again. Reloading can show Retry while the process retains a valid original request. New input, configuration changes, Stop, cache release, or process restart can invalidate it. Historical errors cannot restore a lost request. Original input, tool definitions, configuration, and attachments must still validate.

Answer questions and approve operations

Pending Agent questions, approvals, private inputs, and custom interactions appear in StatusPanel's Interaction category. Click the pill to expand it in either one-to-one or group chats. In the conversation that started a task, you can handle requests from every level of its delegation chain, including work delegated to other Agents. The source shows the task name and executing Agent; select it to view the original execution session. You can respond without opening a child session. Requests from unrelated tasks stay outside this conversation.

When a child Session has a request you can handle, its ancestor conversations and root conversation show Action required in the Works sidebar. If the chain belongs to an Agent's own Chat, that Agent's row shows the same status in the Agents sidebar. This includes multiple delegation levels and work across Agents, even when the child is outside the loaded sidebar pages. Once all requests in the chain are handled, the sidebar resumes showing the conversation's current execution or unread status.

Interaction content appears inside StatusPanel. Use the left and right arrows to switch between pending items; group items identify the requesting member. Switching pending items or leaving the conversation releases the current form, so unsubmitted input must be entered again. Collapsing tool details preserves the active form. For multiple questions, use the left and right arrows to revise an answer. For non-private questions, an optional note is always available as a single-line input beside the Submit or Next button.

Long content scrolls inside the pending area while the chat input stays visible. Inbox and the initiating conversation show the same original request; handling it in either place removes it from pending in the other. Answers and approvals return directly to the original execution session, and approval decisions remain yours. Failed submissions display an error; ordinary answers remain available for retry, while private input must be entered again. Private answers stay out of the initiating conversation's messages and notification records. Requests whose execution was interrupted or whose source is unreachable remain available for status review in Inbox. Tool details show execution information, and pending interactions are handled in StatusPanel's Interaction category.

Compact context manually

In an existing one-to-one Session, type / in the composer and select /compact to fold older stable conversation content into a context summary. The original chat history stays available. You can submit this while the session is running. Compaction joins the Session queue and runs in queue order at a maintenance checkpoint. Compaction and prompts share the Session queue's busy state. The session loading indicator remains active until the queue drains, including across consecutive commands. The chat timeline shows progress and the result, including when there is nothing to compact. New conversation drafts do not show this command.

Tools execute once and save complete canonical results before disclosure. Model-visible business values up to 16 KiB of final UTF-8 JSON retain their structure; larger values return a head-and-tail preview, a Unicode character omission count, and result_ref. The general lookup tool reads that immutable result with opaque cursors, at 100 lines by default and at most 200; each complete packet fits within 16 KiB. History replay and summaries consume only disclosed content and pages already read, without loading unread originals. Binary payloads retain media type and size metadata; original records stay complete. Summaries omit reasoning and UI-only data, retain attachment references, and render user bodies, Agent bodies, and disclosed tool results as ordered plain-text records.

Before each request, the runtime estimates the final system, dynamic summary, messages, and tool definitions, reserving output and safety space. Regular requests allow 8,192 output tokens and summaries allow 4,000; safety space is the larger of 2,048 tokens and 5% of the window. Models must provide context-window metadata. Estimates use roughly three ASCII characters per token, code-point estimates for other characters, and higher density for long encoded strings. They are not actual usage.

Compaction targets roughly the older 65% of stable history. Retaining 10,000 recent tokens is a soft target that request-budget and explicit provider-limit recovery may relax. Each summary selects a complete Part prefix that fits alongside the accumulated summary. Preflight allows at most eight advances and verifies that input shrinks after each one. No progress, oversized fixed system/tools, an oversized user message or tool input, or no eligible stable history produces an explicit diagnostic.

A non-empty summary is saved with its boundary only after normal completion. Failure, cancellation, interruption, or an output-limit finish preserves the previous checkpoint. Summary transport interruption retries at most once. Federation reads at most 16 KiB of provider error text for classification and exposes only a standard error code, status, retry semantics, and an available request ID. Raw error text is not returned. Explicit context_length_exceeded triggers limit recovery; an ordinary 400 remains a request rejection. Recovery rebuilds the next model request with completed tool operations retained in canonical history.

Window and panels

Feature views show the Sidebar Navbar on the left and a content Panel on the right. Settings uses the full Sidebar content width with the Navbar hidden. Click Back at the top of the settings sidebar to return to the page you were viewing before opening settings. A 12px gap separates the back row from the section list. Settings items keep 8px outer spacing and 8px inner padding on both sides. The account and settings menu remains at the bottom in both views. The Sidebar Navbar leaves 8px between the window's left edge and its vertical pill, with spacing following the interface scale. The pill groups compact icon-only entries on a shared background, with names and shortcuts shown on hover; selecting an entry opens the corresponding feature in the content Panel. Switching entries slides a circular selection background along the track and emphasizes the current icon. Initial display and resizing position it immediately; reduced-motion settings disable the animation. The bottom entry spans both the Navbar and content Panel. It shows your avatar and account name when signed in, or a gear icon and Settings when signed out, with an upward arrow on the right. Click it to open an upward menu aligned to the entry's left edge and width. Personal center opens account settings; Settings opens General settings, or preserves your current section if settings are already open.

Feature and settings sidebar lists keep the same available width as content expands or collapses. Scrollbars appearing or disappearing do not shift the list, and header actions stay aligned with the right edge of list rows.

Sidebar Navbar entries show unread dots and the Inbox badge. Hover over an entry to see its name, shortcut, and unread status. The Inbox badge counts pending requests and unread failures; ordinary unread notifications use a dot. See Inbox and unread indicators for details.

Use ⌘/Ctrl + 1 through ⌘/Ctrl + 9 to switch to the first nine Sidebar Navbar entries in their current visible order, from top to bottom. Inbox appears first when available. Number shortcuts follow the visible order as Power entries change. The bottom account and settings menu is outside this numbered sequence.

The Sidebar, main view, and right-hand BayBar share the window. Resizing the window preserves your Sidebar width where space allows, while the main view grows or shrinks first. When space runs short, BayBar narrows and then hides automatically. The Sidebar stays open or closed as you chose. As the window shrinks further, the available space limits the Sidebar width and the main view continues to narrow. Closed panels release their full layout width.

Drag the boundary between a panel and the main view to adjust its width; releasing the boundary saves your choice. Temporarily shrinking the window does not overwrite saved widths. Widening the window restores the Sidebar's saved width, and automatically hidden BayBar returns once there is enough space. Manually closed panels stay closed. Explicitly opening BayBar or following a file link in a narrow window displays it over the right-hand content area. Resizing the window returns it to the available-space layout.

When open, the Sidebar occupies the left layout column and the main view uses the remaining space. Use the top toggle or ⌘/Ctrl+B to open or close it; resizing the window preserves that choice.

Use the top window controls or ⌘/Ctrl+B to toggle the Sidebar and ⌘/Ctrl+L to toggle BayBar. Window resizing and panel dragging follow your movement directly. Panel toggles use a short animation, omitted when reduced motion is enabled.

Work and chat pages

Work Sessions show the Agent avatar and conversation title on the left of the header. Click the avatar to open the Agent profile. The Workspace button sits on the right before the conversation actions menu. Agent replies display their content directly with one shared footer. While the row carries the current execution status, the footer keeps the status and live file changes visible. After execution it switches to copy, more actions, and the timestamp, shown on message hover or keyboard focus. The timestamp follows the more button, and both phases share the same row height and position. The message list determines which row carries the running status in a continuous reply, so an earlier message closing does not produce two competing footer rows. Messages use tighter outer spacing. Markdown in user messages and Agent replies shares the Standard body size (0.875rem) by default and a 1.65 line height, with consistent paragraph spacing. Change the conversation text size in Appearance settings. New work drafts can select a Workspace; saved Sessions offer pinning, sharing, renaming, archiving, and deletion through their actions menu. Chat Sessions use a separate chat page. An Agent's main chat shows its avatar and name in the header, which opens Agent details. Channel chats keep their own conversation title and Workspace context, while groups use the group chat page.

Each page supports message reading, history loading, questions and approvals, drafts, attachments, and reference previews. Drafts and reading state remain scoped to their own Sessions when switching pages, and open BayBar tabs remain available.

Session hierarchy

The Works sidebar’s main list, including the Pinned section, shows only root Sessions. Unpinned roots show running Sessions first, then follow activity time, newest first, with a fixed order for equal timestamps. Pinned roots keep their chosen order. Child Sessions have no separate sidebar entries, including when their parent has not loaded or belongs to another Workspace. Main-list pagination and range selection apply only to root rows.

When a root has loaded child Sessions, a numeric tag sits immediately to the right of its title and counts all loaded descendants. The tag only displays the count and has no click action; it is hidden when there are no children. Expand Sessions in StatusPanel above either the main chat composer or the composer expanded into BayBar. Clicking an entry opens a dedicated BayBar conversation tab; opening the same child Session again reuses its existing tab. The main page keeps the current conversation, and its root row stays highlighted. The child Session tab shows messages, live execution status, and pending interactions, and supports loading message history.

Archiving a parent also archives all descendant Sessions. Finish or stop any running or queued work in the subtree before archiving.

Pin conversations

Pins restore their original conversation record after restart. Archiving or deleting a conversation removes it from active pins; a new conversation with the same name does not inherit the pin. A failed pin read reports its error while other pins continue loading.

In the Works sidebar, open a conversation's more menu and choose Pin to top. Its root conversation moves into a Pinned section above every Workspace and stays visible when Workspaces are collapsed. Pinned roots keep their pin order, newest first. Each root row shows its Workspace and a numeric tag immediately after the title with its loaded descendant count. The chat header menu offers the same action as Unpin. A child’s conversation actions menu shows “Pin parent conversation” or “Unpin parent conversation”; both actions apply to the whole tree. Unpinning returns the root row to its Workspace. New children automatically follow the root. Pinning applies to Works sidebar conversations, including Group conversations.

Reference a Session

In the Works sidebar, open the more menu for an Agent Session and choose Reference. A removable Session · title label appears in the current chat draft. You stay in the same conversation, and the action is disabled when no chat input is available.

You can also choose Copy Session ID in the sidebar or conversation header, then paste into the chat input to insert the same label. Pasting into a plain text field or terminal gives you the original ID. Click the label to see the Session ID and its storage path, which matches Copy path in the menu. Sending the reference shares those details without including the message history.

Share messages and conversations

Shared images and PDFs use 1.5rem padding on all sides. A footer after the messages places the conversation title on the left and the Downcity logo and name on the right. Title and brand text use a small 0.75rem size, with a 1rem logo. Both formats use the chat message components: user messages keep their right-aligned bubbles, background, and corners; Agent messages keep the current page's avatar, name, and spacing between consecutive replies. Timestamps stay in their chat positions. Full message text is expanded, while message toolbars and selection indicators are excluded. Preview and saved files use the same content.

Export captures the current message column width, conversation text size, theme, and interface scale when you click. Ordinary paragraphs keep their chat line breaks. Wide code, tables, and diagrams expand onto the canvas at their original text size and column widths to avoid clipping; other text retains the original message column width. The lower-left image controls offer Standard 1×, High 2×, and Ultra 4×, with High selected by default. Changing the scale regenerates the preview; copying and saving use the selected scale. This changes output pixels without changing message layout. Higher scales usually produce larger files, while embedded image detail remains limited by its source; long content may split into more images. PDF and Markdown do not show this control. PDF scales the same content proportionally to A4 and paginates it, keeping text and vector diagrams as vectors and preserving the proportions of text, bubbles, and spacing.

Choose Share from the more menu on a user message or a completed Agent reply to enter message selection, with the current message selected automatically. Messages keep their original position and width. Circular controls in the left margin and a subtle full-row background show selection. Click message text, whitespace, or the control to toggle selection. During selection, message text cannot be selected, and links, expansion buttons, and other embedded controls do not respond to clicks. Clicking a message only toggles selection; scrolling remains available. Message actions such as editing and forking are hidden. Select additional user messages and completed Agent replies across turns, including non-adjacent messages; exports always follow conversation order. Select all selects eligible messages currently loaded into the conversation. Load earlier history to select more messages; scrolling and loading history preserve your selection.

Click Export in the bottom bar to open the preview, then choose Save as image, Save as PDF, or Save as Markdown. Exports include the sender, timestamp, and full message content. Image and PDF message bodies use the conversation's Markdown typography, text size, and theme, with toolbars hidden. The floating single-row bottom bar matches ChatInput's width, with a background, rounded corners, and space around it. It places Select all on the left, with Cancel and Export on the right. The export button includes the selected count, such as Export (3). With no messages selected, it shows Export and is disabled. Closing the preview returns to your selection; Cancel exits selection mode.

Choose Share from the Chat Session's top-right menu, then Save as image, Save as PDF, or Save as Markdown. Exports read the complete history and preserve the original order of user messages and Agent replies. Reasoning, tool activity, and internal context are excluded. Markdown keeps editable text and attachment links. Images and PDFs preserve code blocks, tables, math, images, and static views of built-in visualizations. Third-party or unavailable visualizations use their published text summary.

The share dialog places its title and conversation name at the upper left, a pill switcher for Image, PDF, and Markdown in the center, and Close on the right. A neutral canvas shows the export directly, without adding image borders, rounded corners, or shadows. Images always scale proportionally to the available canvas width, with small gaps between long-image sections. The 1×, 2×, and 4× settings control output pixels while keeping the displayed preview width unchanged. Detail differences can be subtle at the same preview width; saved pixel dimensions increase with the selected scale. Image quality controls sit at the lower left, while icon-labeled Copy and Save actions sit on the right. PDF and Markdown hide image quality controls. Controls wrap on narrow windows.

Changing format or image quality keeps the previous successful preview visible with a loading indicator. Copy and Save stay disabled until the new preview is ready, preventing actions on stale content. After saving, brief feedback and Open saved files appear while the Save button stays on the right.

PNG shows the generated image sections, with a Copy image button in the bottom bar that copies the section most visible in the preview to the clipboard without saving a file first. PDF shows the paginated file, and Markdown shows the source to be saved. Click the selected format's Save as… button to open the system save dialog. Cancelling that dialog keeps the preview open; closing the preview releases its prepared content. Resource failures appear in the preview.

Short images use a file save dialog. Long content asks for a directory and creates a new folder containing numbered PNG sections. PDFs paginate automatically and keep selectable text. After saving, you can open the file or image folder directly. Cancelling does not create an export. Image resource failures show an error so you can fix the resource and retry.

Conversation text size

Under Settings → Appearance → Conversation text size, use the segmented control to choose Small, Standard, Large, or Extra large. Standard is the default. Changes apply immediately and are saved on this device; select Standard to use the default size.

This setting affects user messages, Agent replies, and composers in Work, Chat, Draft, and Group. Headings and code scale with the body, while the body line height stays 1.65. Message exports follow the conversation text size. Buttons, navigation, file previews, and task documents retain their sizes. Conversation text size and overall interface scale are separate settings.

Chat composers, message rewrite editors, the pending queue, and plain-text message previews share the Markdown body's system font, weight 400, 1.65 line height, and natural letter spacing. Text colors follow the current theme. Adjacent composer paragraphs have the same 0.9em gap as message text; Work, Chat, Draft, and Group use the same text rules. The four conversation text size presets continue to control font size.

Markdown reading

Messages, reasoning and tool details, task content, Workspace READMEs, file and attachment previews, and exports share the same Markdown typography. Body text defaults to Standard (0.875rem) with a 1.65 line height. Chat messages and their exports follow the conversation text size setting; file previews and task documents retain Standard. Headings scale with the body using fixed em ratios, and headings and strong text use weight 500. Text, code, tables, and diagrams follow the current theme.

Standalone documents use a centered body up to 45rem wide, with 2rem vertical and 1.25rem horizontal padding. Embedded content fills its panel without extra outer padding. Both layouts use identical fonts and content spacing. Long code, wide tables, and math scroll horizontally inside their own regions, while the page handles vertical scrolling. Images stay within the reading width, and long links can wrap.

Messages and files

Opening or switching a one-to-one Session positions the message area at the latest message and keeps it at the bottom while the initial messages load, message sections appear, and the viewport height adjusts. This initial positioning applies regardless of the automatic scrolling setting; scrolling upward stops it immediately. After the initial view is ready, new content follows the automatic scrolling setting. Browsing upward preserves your reading position; scroll back to the bottom or choose Jump to latest to resume following.

With fewer than 96 loaded visible messages, all messages stay on the page. Longer conversations recycle offscreen messages while protecting the current text selection, focus, and content being edited. Unmounted history is unavailable to native find across the conversation, and text selections cannot span it. After continuous panel width changes settle, message layouts are remeasured in batches.

Click a workspace file link in a conversation or Markdown document to open its own BayBar file tab while keeping the main page visible. Each Workspace and file path identifies one tab; reopening the same file activates it and updates the line location. Files in other registered workspaces also open in BayBar, and tabs stay available when you switch conversations. Relative links such as ./guide.md and ../README.md resolve from the source document's directory, including documents already open in BayBar. Links with a line number jump to that line. In-page links such as #installation scroll within the current Markdown document. HTTP(S) links open in your default browser.

The file panel and Workspace view select a preview by file type. Markdown renders as a document; local .html and .htm files render as static web pages with relative stylesheets, fonts, and images. Scripts and in-page navigation are disabled. CSV and TSV files display as tables with row and column numbers, preserving quoted separators and line breaks. Tables display up to 1,000 rows and 100 columns. Use Source mode in the file menu to read the complete original text for Markdown, HTML, CSV, or TSV; text links with line numbers open in source mode.

Relative image paths in Markdown files, such as images/chart.svg or ../images/chart.png, resolve from the current document's directory in both BayBar and Workspace previews.

Fenced Markdown code blocks labeled mermaid render as diagrams near the viewport and adapt their layout to the available width. Offscreen diagrams pause layout updates and show the latest content when you return. Diagrams follow the current theme. Copy their source or open the fullscreen view to zoom, pan, fit to the window, and download a PNG with the current theme. Exports include the complete diagram regardless of the viewport. Incomplete streaming syntax keeps the last successful diagram visible; if parsing still fails after generation finishes, the source and a retry button appear.

Fenced Markdown code blocks labeled svg render as images that fit the available width, in both BayBar file previews and chat messages. Incomplete streaming content, invalid SVG, or image load failures display the source instead. SVG loads in an isolated image context without executing scripts.

PNG, JPEG, GIF, WebP, SVG, AVIF, APNG, BMP, and ICO files display as images. Switch between fitting the panel and viewing the original size. Images load separately from the text reader. Other binary files show an Open with default app button that opens the file in its associated system application. If opening fails, the reason appears and you can retry. Text previews remain limited to 2 MB; code and configuration files display as source.

Select text in the right-hand BayBar file panel to show a floating Quote button near the selection. Clicking it adds a reference to the current chat input with the workspace, file path, and selected text; source mode also includes the selected line range. The reference preserves the content at the time of quoting and can be saved with the draft or removed before sending. Clearing the selection, scrolling, switching files, or updating the file hides the button. Markdown, table previews, and source mode support text quoting. Switch HTML previews to source mode to quote text; images do not support text selection references.

One-to-one and group chats share a centered reading column. Messages, the composer, and the context row above it use the same width, with side margins that adapt to the available width. In chat and group pages, a light signature row above the body pairs a 28px circular Agent avatar with a regular-weight 13px name in secondary gray, separated by 10px, with a 10px gap before the body. The timestamp appears when the message is hovered or receives keyboard focus. Click the avatar or name to open Agent details. The top-right of the Chat main view keeps the conversation actions menu; use the identity entry to access Agent settings. Agent content and file changes appear in a left-aligned bubble with a subtle neutral background, a uniform 20px radius, 14px horizontal padding, and 10px vertical padding. Short messages fit their content; long text, code, and tools can use the full reading width. Message actions and runtime status sit below the bubble. Work uses a compact, full-width document without an identity row. In long replies, the identity row stays at the top of the current message while scrolling. User messages align to the right with a slightly deeper neutral gray background, 14px horizontal and 8px vertical padding, and a uniform 20px radius on all four corners for both single-line and multiline messages. User-message bodies, attachments, references, and collapsed previews share compact typography in one-to-one and group chats, with a default body size of 13px that responds to the conversation text-size setting.

Paragraphs, lists, and headings use the shared Markdown reading spacing. Tables use weight-500 headers and subtle dividers, with wide tables scrolling within their own region. Message actions such as copy, quote, and fork appear below all content, including the file changes card when present.

Select text in a user message or Agent body in one-to-one or group chats to show a floating Quote button near the selection. Click it to add the selected content to the current composer. Right-click the selected text to Copy or Quote in composer; the same actions are available in BayBar session panels. When the selection contains HTTP(S) addresses or linked text, Open link opens a single address in your default browser. Multiple addresses appear in a submenu so you can choose one at a time; duplicate addresses appear only once. Messages and BayBar file selections share a pill-shaped button with the segmented switcher's rounded shape and compact sizing, an opaque theme-aware background, a thin border, and hover feedback. It displays only the Quote label. Drag to select, double-click a word, or triple-click a paragraph, including the final paragraph of the body. Each selection must stay within one body block. Triple-clicking a paragraph still allows quoting when its ending boundary passes hidden or unselectable actions. Selections spanning messages or actually selecting action text do not show the button. Clearing the selection, scrolling the message area, or resizing the window hides it.

Within an Agent message, text blocks, reasoning summaries, and tool activities use consistent vertical spacing: 4px at the default zoom level. Paragraphs, lists, and code blocks within the text retain their own reading layout. Collapsed tool activity shows its status and target; expand it to read the available result and any preview limits.

Reasoning, tools, and session activities appear in their original order; consecutive tools can share an activity group. A collapsed reasoning summary uses the last standalone Markdown heading or complete standalone **heading**; otherwise it uses the last non-empty line outside code fences. Inline emphasis does not become a heading. Markdown markers are removed and the summary is limited to 120 characters. Reasoning starts collapsed. Expanded details use the same background, indentation, and border as tool details, with a maximum height of 15rem and scrolling inside the panel. Markdown preserves paragraphs, emphasis, lists, and code. Long reasoning previews show up to 64 Ki characters or 500 lines and indicate omitted content. Adjacent **Heading one****Heading two** titles appear as separate bold paragraphs; code and escaped content retain their source. Its text size follows the conversation setting: Small and Standard use Compact (11px at default scale), Large uses 13px, and Extra large uses 14px. Streaming and completed content share the same typography. Collapsing releases the detail content. Click an activity summary or use the keyboard to expand or collapse its details while the answer stays in place. Write and edit tools automatically expand while receiving streamed input; your manual choice always takes priority.

File reads show only their status and file path, without expandable details. Click a file path in an individual read activity to open it in BayBar. The collapsed activity group summary is plain text. Expanded write activities show their content directly. Markdown uses document formatting and recognized code files use code formatting. Write and edit status appears in the collapsed row. Edit details show the changes, with an explanation only when changes were not saved or the result is unconfirmed. A failed atomic edit applies none of its replacements.

Content searches and file discovery show only their target, result count, and status, without expandable details. Counts refer to returned items, and partial results are indicated in the status row. Write and edit previews show up to 64 Ki characters or 500 lines and indicate omitted content.

Question activities display read-only questions and recorded answers, using option labels for selections. Private answers only show that an answer was provided. Pending forms stay above the composer. Session history copying, context compaction, and command actions use readable localized results; unknown actions keep their name. Activity groups prioritize waiting items, then failures, for file, search, and question tools, and prioritize failed Session Actions before the latest ordinary activity.

Activity details show readable content, execution results, and necessary error messages. Details load only while expanded, and reading a past activity does not execute it again.

Each Power supplies its activity title, target summary, and detail content through its activity renderer. Only actions with details have an expansion arrow; Powers without the slot show a single status row. Collapsed activity group summaries contain no links or interactive controls.

Expanding or collapsing content preserves the current reading position. In one-to-one chats, manual expand or collapse choices for tool/activity groups and long user messages, along with file changes list selections, persist while the current session's local cache exists. They are restored when you scroll away and back or switch away and return. These states are not permanently saved across restarts and are released when the cache is evicted or the session is deleted.

When a skill lookup succeeds and its instructions are displayed, the same call appears once as an expandable entry with the skill name. Failed calls and calls awaiting a response keep their status records.