Chat

How the chat pane behaves: the session rail, the composer, and how the transcript renders.

Tool activity shows a tool-specific icon beside its purpose or details instead of repeating the tool name. When no distinct purpose or detail is available, the row shows the tool label instead of leaving the text blank. Hover the icon to see the exact tool name; screen readers retain that identity. Expanding an activity group keeps the individual tool details and outcomes available. Completed group summaries retain their operation counts. Tool Search calls use the called tool's name, icon, and input details in tool rows and activity summaries. Built-in tools use semantic icons, including an envelope for messages, a calendar clock for scheduled jobs, and an image for image generation. Unknown tools use a puzzle icon unless their plugin supplies an activity icon.

A tool stops showing Running when its completion arrives, even while the parent turn continues. If that completion does not establish success or failure, the row shows Outcome unknown. Partial output alone does not finish a tool.

While a turn is still working and has subagents running, its working indicator ends with their count, such as 3 subagents running, and counts down as they finish. Selecting the count opens the Subagents panel on its list. Child sessions that are not subagents are not part of that count.

Under the reply, each unfinished subagent has its own row showing its name and current activity headline. Rows stay in launch order as activity changes and disappear when the subagent finishes. Until a current-run headline is available, the row shows Running, Queued, or Waiting on subagents when that child has delegated work of its own. Select a row to open that subagent in the Subagents panel, or its session when the panel does not list it. The rows reuse the chat's loaded subagent list; they do not load each child's transcript.

When a turn hands off with sessions_yield and its subagents are still active, their live rows stay under that reply without a separate Waiting on… line or wait timer. The same applies when a turn ends without a handoff. Until the subagent list is available, Waiting on subagents remains as a fallback; its elapsed time counts from the handoff when known. The parent's own working indicator remains visible while the parent is actively working. Child sessions that are not subagents are counted without names once no subagent is left, as Waiting on 2 sessions. Once everything it waited on has finished, the line goes away until the agent resumes. Tool rows you opened stay open through the handoff. A successful sessions_yield leaves no marker in the transcript, and its private continuation context stays hidden.

When the last subagent finishes, the wait line ends and the block stays as it is, without a working indicator, until the turn resumes. Its answer then continues in that same block, with one footer at the end. Tool activity that resumes with nothing written in between joins the activity row from before the handoff. In dashboard sessions, tool activity recorded after the resumed answer, such as the step that sent it, joins the activity before that answer, so the answer stays last; a step that failed there stays where it happened. The working indicator and the closing Done in… line then describe the whole request: time since you asked, including the wait, and output tokens from the runs in that block. The token count is left out when the pane did not see all of those runs, for example after a reload during the wait. When the loaded transcript does not show your request leading straight into that block, for example when the request is older than the loaded history, both lines describe the resumed run alone. A message you send after the handoff starts a block of its own, with that run's own clock and closing line.

In the tool activity, a subagent's launch row shows the label its launch gave it, when it gave one, rather than its instructions, followed by running while it works and by its duration once it has finished. A subagent that failed or timed out reads failed instead, and one that was stopped reads stopped. Selecting the name opens that subagent in the Subagents panel; the rest of the row still expands the launch's details. Collapsed activity counts subagents on their own, such as 1 other operation · 3 subagents. A launch that was refused started no subagent and is counted with the other operations. While the step in progress has no title of its own, the collapsed row keeps showing that count. A launch that opens a child session in its own right, such as one asked for with visible, is not a subagent: its row and its place in that count stay those of an ordinary operation.

The running count, the wait's count and name, and a launch row's state and link come from the session's subagent list. Until the pane has loaded it, the working indicator shows no count and a wait reads Waiting on subagents; a launch row whose subagent is not in the list shows its name alone. The pane does not load that list when Swarm is turned off with tools.swarm: false.

The count and the names lead to the Subagents panel only for subagents that panel lists, including swarm workers, and only in a pane that has the panel. A subagent on the ACP runtime is not in that list, so its name opens its session instead, and a count that includes one is plain text. A pane without a side panel of its own, such as the transcript the Subagents panel shows or a session embedded in Home, opens the subagent's session as well. Waiting on 2 sessions is always plain text.

When your role or session policy blocks messages, the composer is disabled and shows the reason before you try to send. This includes sandbox requirements, shared-session write restrictions, and read-only operator access. Your existing draft stays in place when access changes.

Saved split panes keep their conversation and arrangement. If an older saved pane does not identify its agent, it shows Choose a conversation instead of opening another agent's history. Select that pane, then choose Home or a conversation in the sidebar. Home works even when you are already on its route; the other panes stay unchanged.

When the macOS app embeds a single conversation beside its native sidebar, the chat header fills the window's titlebar row. Drag an empty area of the header to move the window; its buttons keep their normal actions.

Collaborator drafts

In a shared session, another person’s in-progress message stays visible when they pause typing. Draft previews update up to four times per second, keeping the latest text when input arrives faster. After 10 seconds without typing, its label changes from is typing... to Draft without removing the bubble or shifting the transcript. Typing again updates the same bubble. Sending, clearing the draft, leaving the composer, or leaving the session removes it. An idle preview fades out and smoothly closes its space during the final 300 milliseconds before its 30-second expiry, so an abandoned tab cannot leave it visible indefinitely. Typing again cancels the exit. With reduced motion enabled, the preview stays still and disappears at the same deadline without animation. Draft previews are temporary browser state, not saved messages; changing sessions or reconnecting clears them.

Busy sessions keep the first two people’s preview bubbles in arrival order, including their separate Draft lifecycle. Only actively typing people beyond those two appear in the overflow row: its bounded stack shows up to five avatars and one sentence. One, two, or three active overflow collaborators are named naturally; four or five use the first two names and the number of others. With more than five active overflow people, the sentence becomes Several people are typing…. Stopping, pausing, sending, or leaving removes a person from that row immediately; the row disappears when nobody remains active there, even while the first two preview bubbles retain drafts. Long names remain visibly part of the sentence, with responsive ellipsis and full accessible names; enlarged text can wrap. The group reserves two text lines. Its first avatar aligns with the preview avatars, and the sentence follows the actual stack width with a small gap. Active typing text uses a subtle shared shimmer; preview Draft labels, reduced motion, and forced colors retain legible static text. Editing or pausing does not reorder the two preview slots; a slot opens only when its person sends, leaves, clears, or expires.

Reactions

Hover a saved prompt or assistant reply and select Add reaction beside the message actions, or the + chip at the end of an existing reaction row. The quick palette offers 👍, ❤️, 🎉, 👀, 🚀, and 😂; an emoji you already placed shows pressed, and selecting it again removes it. Arrow keys move through the palette and Escape closes it.

Select … for any other emoji. Type or paste one, or open the system emoji picker with ⌃⌘Space on macOS or Win+. on Windows: a complete emoji is applied as soon as it lands in the field, so the system picker needs no extra keystroke. Anything that is not a single emoji stays in the field with a hint. Backspace in an empty field returns to the palette.

Reaction chips appear directly below each message with a count and animate when the count changes. Hover a chip to see who reacted, with you listed first. Select a chip to add your reaction, or select a pressed chip again to remove yours. Viewers can read the chips; adding or toggling requires permission to send or suggest in that session. See Multi-user reactions for agent and channel behavior.

Session rail and side chat

While you watch a running session, the Gateway shows the model's latest safe preamble immediately as the session headline. When a utility model is available, it can replace that headline with a richer compact status digest after enough activity accumulates. Chat carries the result in a session rail: its compact pill shows the live digest, while the expanded rail shows pull requests, elapsed time, and a read-only Side chat thread. The rail can expand once when a run becomes stuck or needs input, and done or failed runs keep a frozen “finished” time based on the final digest. On wide chat panes the expanded rail docks as a 400 px right column; on narrower and mobile layouts it remains an overlay.

Side chat answers questions about the selected session and its project without entering or interrupting the main agent run. On the first question, the Gateway lazily loads a bounded visible snapshot of the selected session before starting the utility model. If history is temporarily unavailable, the question stays visible with Retry instead of being treated as an empty session. Side chat uses read-only access to the target session's history/search and agent workspace. Its bounded thread is held in Gateway memory, is restored when you switch sessions in the Control UI, and is cleared by the rail's trash button, a session reset or deletion, Gateway restart, or idle expiry. It never enters chat.history, and private reference context is not stored as operator dialogue. Open it with Shift-Command-S on Apple platforms or Ctrl-Shift-S elsewhere, or type /btw or /side in the main Control UI composer and press Enter to open the rail and focus its question box. Selecting /btw from the slash menu does the same. Add a question after either command to send it to Side chat; focus moves to its question box when the request finishes. Other clients keep their existing BTW behavior.

When the agent's primary model runs on a CLI runtime such as claude-cli, no utility model is set explicitly, and direct provider authentication is unavailable, Side chat answers through that same CLI runtime. Subscription-only installs need no provider API key. On that route each answer is a tool-free, one-shot side question that uses the bounded session snapshot and observer notes instead of the read-only session tools, and image questions are not supported. Installs with working direct provider authentication keep the direct route, including read-only tools and image support.

Opening Side chat, reopening its panel, or selecting its tab focuses the question box. If you focus another input or keep typing while Side chat loads or answers, that newer input keeps focus.

Confirming text with an input method does not send the Side chat question. Finish composition, then use your configured send shortcut or the Ask button.

Editing a Side chat draft does not interrupt loading its earlier answers. Clear side chat removes the earlier content after the Gateway confirms it; drafts, images, and questions added after the click remain.

The Control UI keeps the latest 24 Side chat turns, including failed questions. Sending a follow-up keeps earlier failures in order; Retry resends that question in place. Failed questions stay in the current pane through a reconnect, but are not persisted across a page reload.

The question box wraps and grows like the main composer; Enter (or your configured send shortcut) asks the question, and Shift+Enter adds a line. Highlighting text in a chat message offers Ask in side chat. Add an optional comment in the selection editor, then save it to stage an editable comment in Side chat. An empty question box starts with a short quote; an existing Side chat draft and the main composer stay unchanged. The full selection and comment accompany that question as bounded, temporary context. After the answer, the restored thread keeps the question text, including its quote, but not the attached context. Reselect the passage for a later question.

If the selection alone exceeds the context limit, Side chat opens without adding the selection and shows a notification. An empty question box still gets a short quote; an existing draft stays intact. If a comment or the combined selections exceed the limit, the editor stays open so you can shorten the comment or adjust the pending selections.

Drop an image onto Side chat or paste one into its question box. You can send it with a written question or on its own. Side chat accepts image attachments, not other file types; its previews never add attachments to the main composer. Image questions require an image-capable Side chat utility model. A text-only model produces a visible error instead of answering without the image; choose an image-capable utility model before retrying. Images are available only for the current question and are not retained in the restored text thread. Reattach an image when asking a later question about it. A failed question keeps its image for Retry while the current pane is retained; reloading the page discards that failed input.

Highlight text and choose Add to chat to attach a comment to the main composer. The optional comment field starts on one line, grows to five lines, then scrolls internally. Confirm or press Enter to save; Shift+Enter adds a line. Saving keeps your existing draft and does not send a message.

Saving leaves a small, filled comment marker beside the selected passage. Click that marker, or the pencil in the comment count's hover preview, to reopen the same editor beside it. Save or Enter saves changes; Cancel or Escape discards the edit; and the trash button deletes the comment. Hover, keyboard-focus, or click the composer's comment count to open its preview. Deleting one comment keeps the remaining list open; Escape, a click outside, or moving the pointer away dismisses it. The count's Remove all comments action clears pending comments in one click and returns focus to the composer without showing a notification. Cleared comments cannot be undone. The clear action appears on hover or keyboard focus and stays visible on touch. Clearing pending comments preserves ordinary attachments, the message draft, and comments already sent in the conversation. Archiving another split pane leaves the current comment editor and keyboard focus in place. Saved comments and their source markers follow the composer's existing draft and queue recovery behavior. When you send, each comment is attached as a text file containing the selection, comment, and source message reference; its draft marker is removed. Hover, keyboard-focus, or tap the sent comment count to read its selection and comment in a compact, scrollable preview. Sent comments remain read-only. Press Escape or tap outside the preview to dismiss it.

The headline owns that run's sidebar subtitle instead of heuristic live activity. It is shared with the official iOS and Android session lists. A final done or failed digest remains visible while the session is unread, then the row returns to its normal work subtitle.

Session observation is enabled by default. Safe preamble headlines do not require a utility model; the utility model only owns richer assessments and terminal summaries. In Settings > Appearance > Sidebar, you can turn observation off gateway-wide, inspect the resolved small model and its provenance, or choose automatic routing, disable utility tasks, or select an explicit agents.defaults.utilityModel. The equivalent config controls are gateway.controlUi.sessionObserver: false and agents.defaults.utilityModel: "".

Session links in messages

Session links in messages open inside the Control UI. This includes agent: keys, root-relative chat URLs, and URLs on the current origin or the Gateway's public origin when its applied configuration is loaded. Hovering a link shows the session card when the session is known locally. Unknown or ambiguous session references remain navigable without a card; links to other origins keep normal browser behavior. Document-relative hrefs are never session links; file references such as src/utils/foo.ts and qa-café/index.md retain workspace file handling, including Unicode names and percent-encoded Markdown link destinations. Explicit Markdown file links also support spaces, emoji, and punctuation in filenames; for example, [Read notes](notes/caf%C3%A9%20note.md) opens the workspace file. Bare CSV filenames in authored links, such as [Read inventory](inventory.csv), and code spans also open the file preview. Plain-text and inline-code file detection stays conservative to avoid turning prose into links.

File links in forwarded messages open against the sending session's workspace, including links inside collapsed update groups and message details. Their filename tabs stay in the current chat pane. Files outside the session root follow that session's file-tool read boundary and also require permission to start a turn in that session. These out-of-root previews are read-only.

In Chat and New Session, provider groups retain their expanded or collapsed state when you close and reopen the same model picker. Closing clears the search text without resetting those toggles. Reloading the page initializes the groups from the selected model again.

While composing text with an input method in model search, Enter, Escape, and arrow keys stay with the input method. They do not select a model, clear the search, or move the highlighted model until composition finishes. Open tooltips also leave Escape to the input method.

When authentication status is available, each provider heading in the chat model picker says how that provider is signed in: API for an API key (or an explicitly selected API-key account), the plan name for a provider with one subscription, and Subscription for a provider with several. With several subscriptions, the heading adds the email of an explicitly selected account when the Gateway supplies it, and the Account rows show each account's email; automatic selection shows no account identity. Hover a truncated heading to read the full text.

Suggested tasks

Suggested task cards offer Start in a new session and Start in a new worktree. Both start the task in the background and keep your current conversation and draft open. The card disappears after the task starts; select the new session in the sidebar when you want to follow its progress. Start in this session runs the task in the current conversation.

Before starting a worktree, OpenClaw checks that the suggested folder is a Git repository with a commit. If it is not, the card keeps the prompt and lets you select a registered project or enter the correct repository path. Select Start in a new worktree again to continue; no child session is started for an invalid source folder.

Composer capability menu

Select + beside the chat composer to open attachments and session capabilities in one menu:

• Skills enables or disables individual skills for this session.
• Connectors enables or disables configured MCP servers for this session. A session tag marks values that differ from the inherited configuration.
• Web search enables or disables managed web search plus native OpenAI and Codex search for this session.
• Manage plugins opens the Plugins page.

These controls are sparse session overrides, like the model and thinking settings in the chat header. A capability with no override inherits the current agent or global configuration, and OpenClaw applies the resolved values when the next run materializes its tools and skills. When overrides are set, open + and select 1 override or N overrides at the bottom of the menu to clear all capability overrides for this session and return to inherited settings.

When tools.web.search.enabled is false, Web search stays off in Chat and New Session. The disabled control explains the global setting. If a session has an older enable override, selecting the control clears that override while search stays off. An explicit session disable remains saved.

For attachments in Chat and New Session, Android Chrome, Edge, Firefox, and Samsung Internet send Take photo straight to the device capture input, without opening a web-camera preview first. Photo and File remain separate choices. Desktop browsers keep the preview, capture, and retake flow; iPhone and iPad Safari keep their single Attach… picker. Embedded native hosts keep their existing behavior. Canceling capture leaves the message draft unchanged.

Video files selected in Chat or New Session show a small local frame preview with a play badge beside the filename. The slot keeps its size while loading. If the browser cannot decode the video promptly, the play icon remains. Removing the attachment releases its preview; generating the preview does not upload the video.

In Connectors, administrators can select Add MCP server… and choose a scope. This session saves the server definition globally but disabled by default, then enables it only for the current session. Everywhere saves the definition enabled globally. Transport, authentication, and other server-definition fields are always global. Session policy can override server enablement and deny individual tools through Tool access.

Tool access lists a connector's tools once a run has discovered them. Before that, it explains why the list is empty rather than reporting zero tools: a newly added server has not connected yet, a connected server has not finished listing its tools, or the runtime catalog predates a config change. Sessions that run on the Codex harness keep their MCP connections inside Codex, so their tools do not appear here.

Capability toggles stay disabled until the Gateway, session, and runtime config are loaded, and read-only operators cannot change them. Adding a server requires administrator access. See Connect MCP servers for the Settings, CLI, and config paths.

Emoji shortcodes

In Chat and New Session, type a colon followed by an emoji name, such as :smi, to see a compact list above the shortcode. The list sizes to its matches and stays inside the viewport. Use the up and down arrows to choose a match, then press Enter or Tab to insert it. You can also click a match. Escape dismisses the suggestions without changing your draft. Selecting an emoji does not send the message.

Typing a recognized complete shortcode, such as :smile:, inserts its Unicode emoji directly into your draft. Code spans and code blocks, URLs, escaped shortcodes, and unknown names stay literal. Existing messages are not rewritten. You can still paste emoji or use your operating system’s emoji keyboard; there is no separate emoji picker in the composer.

JSON in chat

Completed JSON objects and arrays in assistant messages and code fences share a Tree view with expandable nested values and a Raw view of the original source. Copy copies the source in either view, preserving duplicate keys, large numbers, and escape sequences. Raw keeps the usual long-code preview, reveal control, and word wrapping.

Unfinished streaming fences, invalid JSON, and JSON beyond the tree rendering budget stay readable as source. User-message fences and passive previews remain plain code without interactive controls.

Chat behavior

The model picker displays your selected preference, even while a reply is sending, preparing, or streaming. You can change the model and reasoning effort without interrupting that reply. Successfully saved choices apply to ordinary queued messages that have not started; explicit per-message overrides keep their existing precedence. The current turn keeps its execution settings, and its recorded model (including any fallback) does not replace the picker label.

During a model change, model and effort controls wait for the new model’s supported settings. Fast mode and context-window changes remain unavailable during an active reply. Read-only, disconnected, and unavailable-backend protections still apply. Sending does not reload the cached model catalog; catalog loading and refresh feedback stays in the picker.

New Session shows the agent's known default model while the model catalog loads. Model choices are cached in memory for the current connection, agent, session, and account, so reopening a picker or returning to a draft can show them immediately. Catalog and account changes invalidate these copies; reconnecting loads current choices again. Reopening a picker after a reported cooldown expires checks readiness again. A catalog refresh keeps existing controls visible, and the Gateway still validates the model and account when starting a run. Repeated changes while a model lookup is pending are collected into one follow-up lookup for the latest choices.

Loading chat history does not disable the model picker. In Settings > Models, the primary-model picker opens as soon as model rows are known; selecting a new default still waits for configuration readiness and mutation access.

Catalog refreshes update the open conversation's model and context facts without reloading unrelated session lists. The shared session store applies lifecycle row snapshots to existing active members locally. Membership or configuration changes and events without a row snapshot refresh the affected lists through its paced event scheduler.

Chat refreshes its available commands after skill selections or session settings change. Repeated changes share one pending refresh per conversation and connection; if a read is already running, one follow-up read picks up the latest changes. Older results cannot replace the current command list.

If a New Session model lookup does not finish within 30 seconds, the controls show Models unavailable. Open the model picker to retry; your draft stays in place. A retry waits for the earlier lookup to finish before starting more work, and its 30-second deadline includes that wait. Reconnecting clears pending lookups from the previous connection.

When you open an existing session, the conversation appears before supporting panels and pull-request details load. You can start typing as soon as its identity is resolved, while the transcript still shows its loading skeleton. The same composer keeps your draft and focus when the conversation appears. You can send ordinary messages and attachments while history loads: the message enters the outbox immediately, leaving the composer ready for your next draft. The open chat confirms its current session and conversation branch before delivery continues automatically. Switching chats keeps queued messages tied to their original conversation. If history fails to load, the queued message stays available while you resolve the history error. Goals and other slash commands wait for history; /stop and /approve remain available. The initial task progress read reserves only its card slot; the transcript and composer stay available.

When you open a short chat link, identity prepared during the current connection can make the composer ready sooner. The original link stays in place until the session lookup confirms the same conversation and its current title.

Other conversations warm only when you hover or focus their navigation entry, after the visible transcript is ready. Simply connecting does not fetch unopened conversations.

Background refreshes for saved sidebar filters, groups, automation status, and the Inbox wait until the conversation appears. Task suggestions and the progress card then refresh after the transcript paints. Changing a filter or opening a group-targeted New Session remains immediate.

Panes share outbox recovery for the same conversation. Activity in another conversation does not restart that recovery; reconnecting checks every saved outbox.

Files remain in the composer while a submitted message waits for attachment storage. If you begin another draft during that wait, saving the first message removes its unchanged files unless another pending submission still owns them. Your newer text and added or replaced files stay in the composer. A storage failure keeps those files available for retry.

On wide desktop panes, a compact rail of horizontal marks sits in the transcript's left gutter only when the conversation exceeds the visible height. It gently fades in as the conversation grows and hides again if the whole conversation fits after a resize; reduced motion disables the fade. Hover for a short message preview, or click a mark to jump to that message. Tab focuses the rail; arrow keys move between marks, Enter or Space jumps, Home and End select the endpoints, and Escape dismisses the preview. At rest, all marks are identical 8 × 2px strokes at 12px spacing. They stay faint; marks for messages currently visible in the transcript light up together as you scroll. Hovering a mark grows it to 32px and lights only that mark in text color, with progressively shorter strokes across three neighbors on either side. The wave and highlighted mark stay in place while the pointer moves onto the preview card. Leaving the rail and card, or pressing Escape, clears the wave. The other marks keep their resting colors. Outside that hover range, widths stay fixed. An empty message preview shows “Preview unavailable.” Each visible user message has a mark, and assistant messages from the same run share one mark. Tool calls, results, and progress alone do not create marks. An assistant mark jumps to the first currently displayed response in its run and previews the latest displayed response. Its identity stays stable as streaming output becomes persisted history, and its current-position highlight follows later response content in the same run. The rail covers loaded history; messages without run identity retain their transcript grouping. Long rails scroll internally within 45% of the viewport height, with fades only at ends that hide more messages. Scrolling the transcript keeps the current mark visible; you can also scroll the rail to explore other messages. The rail stays hidden on mobile, in narrow or short panes, and when your saved message width leaves too little gutter space. A jump briefly tints the target message with a soft background, fading over 1.2 seconds without a border or ring. Reduced motion disables mark transitions and shows the target tint statically for one second.

Session dashboards follow the selected conversation's agent, including when multiple agents each use a global session. Split panes keep their owners separate; panes showing the same agent and conversation share dashboard updates.

Automatic session titles describe the topic or intended task in your first message. The naming model is asked for a plain-text label of that message, not of injected harness or project instructions, and not a copy of the input's JSON wrapper. They are generated separately from the agent's work, so a title is not a completion status or a report of tool access. Existing titles and manual names are left unchanged; click a title to rename it.

Worktree creation waits up to 30 seconds for a title, then proceeds while naming finishes in the background. A late title still updates the session without renaming its existing Git branch. Concurrent naming requests share the same work; if that request fails, a waiting dashboard request retries once. Naming starts when the first reply begins; if it fails while that reply is still running, as on a model server that handles one request at a time, it retries once after the turn ends. If both model routes fail, the session uses a two-word crustacean-themed name. The Gateway logs a warning when chat naming uses this fallback.

Collapsed tool rows keep the tool label visible and truncate long summaries with an ellipsis. Completed answers stay visible outside collapsed work, including when a later Gateway notice arrives. The completed response footer keeps the final answer's timestamp when earlier tool activity is restored after a reload. Expand a tool row to inspect its command, path, or query. Inspect subagent status from the parent conversation with /subagents list and /subagents info <id|#>, or read recent child messages with /subagents log <id|#>. See Sub-agent slash command.

While the agent works, its latest tool-activity disclosure shows the current operation's purpose instead of operation counts. Within that disclosure, each title gets three seconds to read before a newer title replaces it; rapid calls keep only the newest pending update. The last purpose stays visible between tools, without a running ellipsis after that operation ends. The row stays expandable throughout. Failed or blocked operations and the end of the run bypass the hold, and settled activity returns to counts, or to a lone workflow's own row as described below. A new disclosure after inline narration starts with its own current purpose rather than moving an older operation across the narration. Reduced motion disables the title transition.

Tool activity summaries count the operations inside a workflow rather than counting its wrapper again. A workflow whose recorded calls are all routine, such as plan or progress updates, is counted itself instead. Once it is no longer the live activity, activity that consists of one such successfully completed workflow shows that workflow's own titled row rather than a count; expand it for the routine calls and Tool input. A workflow that failed, was blocked, has no recorded outcome, or went through an approval review keeps the counted row with its status. Execution calls show the agent-provided purpose when available; titles describe intended work, while results determine success or failure. Recorded child calls appear under their operation instead of as separate peer rows. Expand the operation to inspect its children, then expand a child for its command, full output, and reported exit status. Tool input retains the wrapper's source and output. Collapsed operations include failures from their children, even when the wrapper or later calls succeed. Error messages and diagnostic paths stay inside the expandable tool details. Nested relationships use recorded call metadata from the same run and survive reloading; calls without an available, unambiguous parent stay separate. Untitled command previews flatten line breaks and truncate long commands; expanded details retain the original source.

Native Codex Code Mode calls show run JavaScript when no purpose is available. Expand Tool input to read the source. Captured text-block responses display their text directly, and completed command envelopes show readable output with nonzero exit codes kept visible. JSON output is indented without changing number or string values. Raw details retains the original response, including execution metadata. For long results, choose Show full output to inspect the complete response; copy and download preserve those captured bytes.

Filesystem paths remain readable in tool activity and error messages; credential values are still masked. Compact tool labels shorten macOS, Linux, and Windows home-directory prefixes to ~ while retaining the directory and filename.

A turn that fails before producing any reply leaves a durable notice in the thread. Failed and timed-out turns also show the available failure reason in the sidebar's compact summary and run-error tooltip, including while a session refresh is still catching up.

Chat error banners, including cloud runner failures, show short messages in full. Use Copy error beside Details in the header to copy the complete diagnostic received by the UI, even while collapsed. Details appears only when the complete diagnostic adds information beyond the preview, such as additional lines or text shortened for the preview; repeated lines and whitespace-only differences do not add details. Open it to read and select the complete diagnostic. The disclosure works with Enter or Space; the expanded text wraps long lines and can be scrolled with the keyboard. Copying does not open or close the details, and neither copying nor expanding an error retries the failed operation. Retry and other recovery actions remain separate from the disclosure.

When an active run compacts the conversation while your next message is being prepared, OpenClaw follows the verified continuation automatically, even if the earlier run finishes before preparation does. This keeps the original message and send identity without displaying a retry error.

Run-error banners offer Refresh to reload the conversation without resending a message or replacing your draft. If the conversation changes before a message can run and dispatch cannot verify a safe continuation, the banner explains that the message did not run and asks you to refresh before sending it again. The original diagnostic remains under Details. OpenClaw does not automatically redirect that message into a replacement conversation.

ClawHub recommendation cards

Ask to find or install a plugin or skill, such as “Find the WhatsApp plugin”, to let the agent search ClawHub. The agent uses available tools and skills first; it suggests cards for explicit discovery or installation requests, or when a needed capability is missing. Routine tasks, tool errors, and permission fixes do not call for a catalog search. When the message tool is available, it can present up to three matching cards in the conversation.

If you use the coding tool profile, include "message" in tools.alsoAllow (for example, tools: { profile: "coding", alsoAllow: ["message"] }). Existing deny rules still apply. See Tool access configuration.

Select a card to open its listing inside the Control UI: plugins open in Plugins, and skills open in Skills. A card's Install button opens the capability details. Select Install on a plugin overview to start installation immediately. Configured install-policy warnings still require an explicit acknowledgment; see Manage plugins. Dismiss dismisses the card from the current view.

An installed capability shows a green checkmark and Installed. This means the plugin package or linked skill is present. A plugin may still need to be enabled, configured, or connected to an account before the agent can use it. The card checks the current installation status; select Status unavailable · Retry if that check fails.

The agent requests cards through message with a capability query:

{
  "action": "send",
  "clawhub": { "query": "whatsapp", "kind": "plugin" }
}

query is required; kind can be plugin or skill. Omit kind to check plugins first, then skills if no official plugin matches. Omit channel and target to reply to the current Control UI conversation. ClawHub supplies the official designation, and the Gateway checks installation status; the agent cannot assign those badges. A search with no official match or an unavailable catalog returns an explanation in chat.

Source previews and copying code

Long clipboard text appears as a file-sized card in the composer and a compact chip in the transcript. Its label shows the first 30 characters of a plain-text excerpt, with HTML and Markdown formatting removed. Empty or unavailable excerpts show Pasted text. In the transcript, chips sit above the text bubble alongside other attachments; multiple chips share a row and wrap when needed. Click the composer excerpt or icon, or a transcript chip, to open the existing attachment side panel and copy the original text, preserving markup, line breaks, and indentation. The excerpt and chip also support keyboard activation. Select Show in text field on the second row inside the composer card to return its text to the draft without opening the side panel. Use the X at the card's top-right corner to discard the pasted text without changing your draft or other attachments. The composer side panel also offers the same action and removal. Messages containing only comment or pasted-text chips use a transparent shell. Newly uploaded text files remain file cards, even when their names resemble pasted-text attachments. Older history without origin metadata recognizes text/plain attachments named pasted-text-<digits>.txt as pasted text.

Uploaded attachments keep their original filenames on download, including spaces and Unicode characters. Uploaded images also retain their names in accessible transcript image and image-viewer labels after sending and reloading; descriptive alt text takes precedence when provided. Select Open on an uploaded text attachment to read it directly in the Files side panel. Plain-text attachments, CSV, and JSON preserve line breaks and indentation. Markdown attachments render as documents with interactive code blocks. When an open attachment refreshes with unchanged text, its code blocks keep your expansion and wrapping choices after loading. A different attachment or changed text starts with fresh controls. Long previews scroll inside their file tab; switching tabs preserves your reading position. HTML files (.html, .htm, or text/html, including MIME parameters) render as a sandboxed page by default. Source and Preview share the same filename tab and slim toolbar. Same-origin HTML attachment previews accept UTF-8 content up to 2 MiB; other text attachment previews remain limited to 256 KiB. Unsupported, external, oversized, or unavailable attachments keep their original Download action. Cross-origin HTML attachments are download-only.

Self-contained HTML can use inline CSS and JavaScript under the default scripts sandbox setting. The page runs on the existing separate-origin sandbox host, not in the Control UI document. It receives no agent-prompt, tools, or dashboard APIs. The global strict setting disables document JavaScript while retaining the same sandbox resource policy and descendant-frame restrictions; trusted does not give ordinary HTML files access to the Control UI origin. External resources remain subject to the sandbox's default content security policy.

Session-file HTML previews load relative images, media, stylesheets, and scripts from the file's folder through the same session read boundary. This also applies to unsaved HTML drafts. Each asset request accepts at most 64 references, with a 1 MiB limit per asset and a 4 MiB total before base64 encoding. Stylesheets get one additional asset request for relative CSS url(...) resources; inline styles use the same resource handling. Cloud repository assets retain their existing 256 KiB preview limit; stopped repositories expose only retained changed artifacts. Fonts and CSS @import are left untouched, so local fonts are not loaded. Documents with an authored <base href> keep all their original references. Missing, denied, unsupported, or oversized assets remain unchanged and produce a compact notice in the preview toolbar. Attachment HTML previews have no session folder and do not load relative assets. Loaded classic scripts with defer run after parsing, in document order with non-async module scripts, while retaining classic script globals.

Authored in-page HTML links such as href="#section" scroll within the preview, including in strict mode. This leaves the original file and Source unchanged; authored base URLs, links targeting another window, and download links keep their behavior.

Workspace file reads and edits retain their existing 256 KiB limit. After a successful save, Files refreshes its metadata and Review refreshes its diff without switching the active panel or clearing the current Files search. Code previews and editors are named by their filename and reachable with Tab, including read-only previews. In-file search announces its result count. For workspace HTML, Edit, in-file search, and explicit file:line navigation open Source. Switching back to Preview renders the latest unsaved editor text without saving it. Switching views preserves the editor, undo history, and draft; switching file tabs preserves each file's mode and unchanged page state. A reopened unsaved draft can be saved directly from Preview, including the same conflict detection and Reload/Overwrite actions as Source. A preview transport failure shows Retry without retrying in a loop, and Source remains available, including when the optional preview cannot load.

View Raw Text opens a Source view that keeps Markdown notation literal, including nested code fences. The raw-text action disappears while that view is open. Decoded text artifacts use the same literal preview. Copy code preserves the code's leading whitespace and final newline when present. Indented Markdown code blocks also work at the start of a message and remain literal while streaming, including blank lines within the block.

Code blocks keep your expansion and wrapping choices when their closing fence arrives and later paragraphs stream into the same assistant reply. Replacing the message, correcting earlier content, or changing rendering options starts a fresh view. References that change earlier Markdown can also reset the view. Completed lists also stay cached as later blocks arrive; loose or nested list continuations remain together until the list ends.

Browser results for X and Twitter posts use a wider card with the author, handle, available post text, and preview image. Open post and More actions stay visible on touch screens. Text and social images reuse the anonymous page preview; when metadata is unavailable, the card keeps the post link and handle. It does not load an X embed script or require an X account.

Copy URL in browser tab cards also works on plain HTTP connections where the browser does not provide its Clipboard API.

If the browser rejects a text clipboard write, starting another text or image copy cancels its delayed fallback. Code blocks replaced during streaming and Mermaid diagrams whose source changes also cancel that fallback. This does not cancel native clipboard writes that the browser has already accepted.

Markdown tables

Markdown tables wrap headings and cell text to fit the conversation. On wide desktop panes, top-level assistant tables stay at the reading width when their content fits and use extra space only as needed, without widening the surrounding prose. Long cells wrap within the pane limit; genuinely dense tables still scroll horizontally. Short values such as byte counts stay intact beside long filenames or hashes, including on mobile. Table controls appear below the table. Wide desktop tables use compact icon-only controls; mobile and phone-landscape views retain larger touch controls and a visible Expand table label. Ordinary inline tables grow vertically instead of adding a vertical scrollbar. Copy table copies tab-separated cells, and Expand table opens a larger view with a sticky header. If copying fails, the button clears any earlier success checkmark. In Chat, workspace file and session links work in either view, including Enter and Space keyboard activation. Following a link closes the expanded view so you can use its destination.

The Ask OpenClaw panel supports table scrolling, copying, expansion, and web links. Its session links open with a click or Enter. Space does not activate those links, and workspace-file references do not open a preview.

Mermaid diagrams

Use a mermaid fenced code block in chat. The Control UI renders it as a diagram automatically:

```mermaid
flowchart LR
  Gateway --> Worker
  Worker --> State
```

Open the Diagram options menu in the top-right corner to switch between the diagram and source or choose Expand diagram for the image viewer with zoom. The copy button appears on hover or keyboard focus and stays visible on touch screens. It copies the original Mermaid text. Diagram colors and fonts follow the current UI theme.

An unfinished streaming fence stays readable as code. Rendering starts when the closing fence arrives or the response finishes. Invalid or overly complex diagrams keep their source visible with an error; correct the syntax or simplify the diagram. Diagram source cannot run scripts or click handlers, load external images, or add custom CSS to the Control UI.

Renderer loading or timeout errors instead suggest reloading the dashboard and checking proxy authentication. The renderer runs in an isolated frame, so its assets/mermaid.min-*.js and assets/frame-*.js requests do not send SameSite=Lax or SameSite=Strict cookies. Behind a cookie-authenticated reverse proxy, those static asset URLs must be reachable without those cookies, including under any configured gateway.controlUi.basePath. Check the browser Network panel for blocked requests or redirects to a login page. Keep authentication on the dashboard and Gateway APIs; any proxy exception should cover only these static renderer assets. Reload after correcting the asset access rules.

Hosted embeds

Assistant messages can render hosted web content inline with the [embed ...] shortcode. The iframe sandbox policy is controlled by gateway.controlUi.embedSandbox:

strict

Disables script execution inside hosted embeds.

scripts (default)

Allows interactive embeds while keeping origin isolation; usually enough for self-contained browser games/widgets.

trusted

Adds allow-same-origin on top of allow-scripts for same-site documents that intentionally need stronger privileges.

{
  gateway: {
    controlUi: {
      embedSandbox: "scripts",
    },
  },
}

Absolute external http(s) embed URLs stay blocked by default, except for the dedicated YouTube video card. To let [embed url="https://..."] load other third-party pages, set gateway.controlUi.allowExternalEmbedUrls: true.

Widgets created by show_widget load through the authenticated Gateway connection in every sandbox mode, including while settings are loading. In strict mode, their content remains visible but scripted interactions are disabled.

When a live widget moves into the assistant's final reply, Chrome and Edge 133+ and Firefox 144+ preserve its playback and local interaction state. Browsers without atomic DOM moves, including Safari, reload the widget during that move.

While a widget loads, a subtle shimmer reserves its space without displaying loading text. With reduced motion enabled, the placeholder stays still.

The core show_widget tool renders self-contained SVG or HTML directly from a tool call. The browser and supported native chat clients advertise the inline-widgets Gateway capability, and the resulting Canvas document remains available when chat history reloads. Channel plugins such as Discord Activities can register contextual presenters behind that same tool. Channel-originated runs without an eligible presenter or inline client do not receive it.

YouTube videos

An assistant can embed a YouTube video directly in its reply:

[embed url="https://www.youtube.com/watch?v=VIDEO_ID" title="Trailer" /]

The Control UI shows a thumbnail and Play button. The player loads only after the user presses Play; no HTML widget or media download is needed. An Open on YouTube link remains available when a video cannot play inline.

Supported HTTPS links include youtube.com/watch?v=..., youtu.be/..., YouTube shorts, live, and embed URLs, and youtube-nocookie.com/embed/.... A start time in start, t, or #t is preserved; unrelated URL parameters are discarded.

The dedicated card recognizes YouTube video URLs independently of gateway.controlUi.allowExternalEmbedUrls. It does not enable arbitrary nested iframes in show_widget. With gateway.controlUi.embedSandbox: "strict", the card offers the watch link without loading a player. Other chat surfaces should use a regular YouTube link; the card is a Control UI feature, separate from native inline widgets.

Thumbnails load from i.ytimg.com, which contacts Google before playback. After activation, the player loads from www.youtube-nocookie.com. Its frame sends only the Control UI origin as the referrer, never the chat URL or query string; YouTube requires that origin to permit embedded playback. Direct video attachments and HTML widget media keep their existing behavior.

Session Details

Select Details at the top-right of the conversation to open the session’s compact information surface. It opens only on request, not when progress or background activity updates. Session details and Task progress have independent disclosure states. Session context groups workspace information, pull requests, people, subagents, and attached automations; it reuses their existing actions rather than duplicating their data. The former pull-request and task-progress bars are not repeated above the composer.

The expanded progress card’s options include Don’t show task progress again. This turns off the existing browser-local Show task progress cards preference and offers Undo. It does not stop work or delete saved progress. Session details remain available, as do dashboard progress widgets and session hover previews. Re-enable cards from Settings → Appearance → Chat.

Chat transcript layout

Suggested-task cards show their short summaries in full. Unusually long summaries scroll within the card, and Show instructions expands the complete task prompt. In short windows, the card body scrolls while the header and Start controls stay visible.

Use the mouse wheel or trackpad over the composer or its surrounding space to scroll the conversation while the footer stays pinned. Persistent notices and the composer reserve their own space below the transcript; session context and task progress are available in Details. Long drafts, Details sections, and menus keep their own scrolling when their content overflows. At the footer’s edge, the wheel continues into the conversation. On touch devices, swipe the transcript or footer to scroll that region.

Scrolling down at the end keeps the final reply above the composer as images or other message content finish resizing. Scrolling upward keeps your reading position instead.

Task progress cards are enabled by default. Toggle Show task progress cards in Settings → Appearance → Chat to hide or show the card in Details in this browser. Hiding it does not stop agent work or clear saved progress.

Opening Details does not couple the two sections’ disclosure states. Sending messages, run completion, and progress updates do not automatically open Details or reopen a card the user collapsed. See Task progress cards for card-lifetime and preference behavior.

Streaming output and layout adjustments keep reading mode intact. While you are at or near the end, new messages and replies keep the transcript pinned to the latest content, including turns started from another browser, device, channel, or automation. Typing previews preserve this follow state. Scrolling up pauses following and preserves your reading position as incoming content grows. Scroll back to the end, select Latest, or send a message from this pane to resume following. Assistant text stays visible as it streams and becomes saved history, without a reply entry fade or slide. Submitted prompts slide upward once without fading out; the smooth send scroll starts after the composer and new rows have settled their layout. Reduced motion disables the prompt slide and smooth scrolling.

Hover an external web link, or focus it with the keyboard, to see its page title, description, and social image when available. GitHub repository and public landing-page links use the same card; issue and pull-request links keep their dedicated previews. The card uses OpenClaw's theme and includes Open in your browser. Press Escape to dismiss it. Pages without metadata keep the link label and destination. Touch taps open links normally. GitHub sign-in, account, and profile links do not trigger page previews. These anonymous previews respect Automatically Fetch Link Favicons; see the request and privacy details.

Completed replies can show a compact Sources strip when their web links match recorded web_search or web_fetch results from the same run. Select a title and domain card to preview its Search snippet or Page excerpt, then use Open source to open the recorded page. Previews reuse retrieved content without fetching pages again. Site icons use the existing Automatically Fetch Link Favicons preference and Gateway icon loader, with a globe when disabled or unavailable. Session links and GitHub issue or pull-request links keep their dedicated previews instead of appearing again in the source strip. Citation-only results keep a link and show No excerpt available. Uncited results, failed tools, and results from other runs are not included. Source cards require run identity in the transcript and follow tool-activity visibility. They are hidden while a run is active or transcript search is filtering messages.

In completed dashboard turns, commentary, reasoning-only messages, and tool activity share one Worked for… disclosure above the answers. Expanding it shows the activity in its original order; explicit answer segments and visual results stay visible below it. Browser tab previews keep their execution order relative to those visible results, so a page opened after an inline widget appears after that widget. Failed tool results after the last answer stay visible outside the disclosure until a later answer follows them. This is display grouping, not a change to stored history. Live turns, search results, and turns without an answer stay expanded. So does a turn that handed off with sessions_yield, whether it is waiting, has resumed, or never did: its activity stays in place, and once it resumes the closing line reports the request. User messages, forwarded inputs, and structural markers remain boundaries for grouping. Selecting the author's name on a reply's Replying to line scrolls to the original message and briefly highlights it, first opening its containing Worked for… group when the original sits inside one.

On wide desktop panes, the conversation position rail provides keyboard shortcuts to messages. In right-to-left interface languages, the rail uses the right gutter and its previews open toward the conversation. Tab enters at the current message, or the first marker if no message is current. ArrowUp and ArrowDown move focus; Home and End go to the first and last markers. Enter or Space jumps to the focused message. Tab or Shift+Tab leaves the rail in one step, and Escape closes the preview and returns focus to the transcript. In split view, Escape handles the focused pane before a hovered preview in another pane. Focusing a marker also shows its preview without jumping to the message. Changing the draft height or entering Goal mode keeps the rail's scroll position stable. Navigation and keyboard focus still reveal the selected marker.

The chat transcript uses a centered readable frame aligned with the composer. Assistant and tool output stay left-aligned while your own messages stay right-aligned inside that frame. In multi-user sessions (for example a group chat relayed from a channel plugin), messages from other attributed participants render left-aligned with the author's avatar, name, and a stable per-identity color, so only the signed-in viewer's messages read as "mine". A reply can carry one "Replying to" line with the author's avatar and name, and no message excerpt. Assistant replies show it when they answer an older message, or when the conversation is shared: two or more people appear in the session participants or anywhere in the transcript, including messages hidden by search or not yet loaded. A reply in a one-to-one conversation that answers the prompt of its own turn shows no line. The line appears only once its origin and author are known, from a loaded message, a fetched original, or a reply snapshot that names its sender; until then, or when the author stays unknown, it is omitted rather than guessed. A reply to the current message resolves only through the run that prompt started, so legacy, mirrored, or ambiguous replies without that link show no line. When the original is confirmed deleted or inaccessible, the line shows "Original message unavailable" without an avatar or link, after the sender's name when the snapshot names it. A line whose row is reserved while the original is fetched never changes height: it fills in place, and a lookup that cannot reach the Gateway keeps the row reserved until a new connection retries. One exception is a shared conversation where the fetched original has no sender metadata and no reply snapshot names its sender: the author stays unknown, so the line ends hidden and its reserved row is released when the answer arrives. On desktop, the agent avatar stays at the top of attributed reply groups, with a short curved connector running from behind it to that line. On narrow screens, agent and participant avatars beside messages stay hidden to preserve content width; a small reply arrow introduces the line instead of the connector, while the author's avatar remains visible. Your own replies keep the line inside the bubble. Select the author to scroll to and briefly highlight the original. System entries such as local slash-command output render as centered notice rows without an avatar.

Images and video previews in your own messages appear above any accompanying text, without a surrounding bubble background. Videos use a still frame with a play icon; select the preview to open the video in the Files panel. If a preview cannot load, the attachment card remains available. Hovering media leaves that layout unchanged, and the text keeps its normal bubble color, including any per-identity tint. Assistant videos retain their inline player. Expand a video to sweep through assistant videos in the same turn with the previous/next buttons or Left/Right arrow keys. On touch screens, swipe the video picture or surrounding space left/right; the bottom native control strip remains available for seeking. When the video player itself has keyboard focus, its native playback shortcuts take precedence. Navigation stops the previous clip, and closing the viewer stops playback and returns focus to the original control. Videos from another turn or session are not included.

Images use a plain, lightly tinted shimmer while their availability or preview is loading. An image that needs explicit permission keeps a compact Allow image card until you choose to allow it; unavailable images retain their status and retry controls.

Managed image previews retain enough detail for high-density displays. Open an image tile to inspect it immediately in the image viewer; the cached preview stays visible while the original loads and is replaced as soon as the original is ready. Reopening an image reuses the cached original. If loading fails, the preview remains visible. When the message contains multiple images, use Previous image / Next image, the left and right arrow keys, or a horizontal swipe to move through that message's images. Navigation stops at either end and reverses direction in right-to-left layouts. Pinch or use the zoom controls to inspect details; dragging or holding Shift while pressing an arrow key pans the zoomed image. The keyboard shortcuts dialog lists these controls. Press Escape, select Close image preview, or click outside the image to close and return focus to the tile you opened.

Inline audio players support five-second seeking with the arrow keys, including when the seek slider has focus. The slider announces the current time and duration.

Images attached to assistant progress messages appear inline while the task continues and remain visible after reloading the conversation. Progress attachments are preserved for runs started by any Gateway owner, including chat sends, restart recovery, and internal dispatch. Remote attachment URLs do not need a filename extension: the Gateway detects the media type and serves the preview through the same authenticated media path used for final replies. Documents keep their file cards.

Computer screenshots retained in the session transcript also load as image previews, including after a reload. Opening a preview shows the saved screenshot in the image viewer. The Control UI retrieves the pixels through authenticated artifact access; computer observations are not automatically sent as outgoing attachments. Historical entries that contain only omission metadata keep the Omitted from history card.

In automatic visible-reply mode, this includes standalone MEDIA: lines in model-authored commentary committed to the transcript, not just final replies. Only references captured before transcript hooks and retained in that commentary are eligible; hook-added references remain text, and normal media access and live run/session checks still apply. Message-tool-only delivery uses message(action=send) with structured attachment fields instead. Tool/plugin output and streamed block payloads must also use structured fields. See WebChat commentary compatibility.

Messages forwarded by sessions_send render as left-aligned speech bubbles with a From attribution row above the message. Known senders, including the current agent, retain their agent identity. Unknown or unlisted senders show no avatar beside the bubble or in the attribution row, and no empty inline avatar space remains. The message column stays aligned with neighboring messages. Select a linked source to open its session; hover it to see session progress. Each source session has a stable bubble tint. Forwarded messages without a known source session show the source agent when available, or a generic forwarded-message label. The receiving agent's own replies remain flat text.

Your name is hidden beneath your own messages when no other human participant is known in the session. Shared conversations keep sender names, including while searching the transcript. Web messages do not show a "via Web" label; other recorded client sources remain visible.

Run transcripts

Open a session to inspect its recorded messages and tool calls.

Automation run transcripts stay tied to the recorded run, including after its temporary continuation session is removed or the automation runs again. If that recorded transcript is unavailable, the viewer reports an error instead of showing a newer run. Select View transcript under Automations → Run history to read that exact run without opening its temporary session. Selecting All automations, opening another automation, or starting a new automation closes the run viewer; refreshing the same panel keeps it open.

The automation run viewer displays the supplied transcript. Core session transcripts are currently capped at 8,000 characters per text block; loading earlier messages does not recover a capped reply's missing text.

Conversations stopped for review

When a provider stops a conversation as a misalignment precaution, chat holds ordinary sends, queued input, and Talk. Review findings opens the available explanation. If the provider and runtime support continuation, the dialog shows the exact continuation message and asks you to Acknowledge findings and continue. Sending that request keeps the chat paused until the provider accepts it; a refresh, another session, or newer findings cannot confirm an older review.

Queued messages remain held after continuation. Review and retry each one separately if it is still needed. Without a supported continuation, the conversation remains stopped. See misalignment precautions for runtime support and recovery limits.

Chat message width

Wide-monitor users can override the transcript width under Settings → Appearance → Chat → Message width. The preference stays in that browser's local storage. Supported forms include plain lengths and percentages such as 960px or 82%, plus constrained min(...), max(...), clamp(...), calc(...), and fit-content(...) width expressions supported by your browser. Invalid input shows an error and keeps the last saved width. Clear the field to restore the default.