Team Chat (built-in channels and direct messages)
Team Chat is Omniscio's built-in team messenger — channels, direct messages, threads, reactions, pins, search and file sharing, delivered live to the desktop app and your phone, plus self-serve workspaces you create yourself and share by invite code. This page covers what it does, where to turn it on, how it behaves, and the engineering behind it, across a parent page and four more parts.
On this page
What it is
Team Chat is a built-in, real-time team messenger for your company — channels and direct messages — delivered live to every device (the desktop app and your phone) via the cloud. It is a lightweight Slack replacement that lives inside Omniscio and uses the Global Auth identity you already sign in with, so there is no separate account to create.
It renders as a native built-in virtual project in the projects sidebar (gated). The channel/DM list sits in the sidebar pane; the open conversation (messages + composer) fills the main panel. It also hosts sessions: the Sessions tab at the top of the sidebar is the standard session list, so you can start an AI session without leaving the chat.
Where to find it
Status: In-development (ships dark). Reveal via Settings → Lab → Team Chat, or launch with AMC_SHOW_TEAM_CHAT=1. Until enabled it is completely hidden — no sidebar row, no setting effect. Omniscio's own developers — anyone running the app from source on their own install — get it switched on once automatically, so the team channel is never missed; switching it off afterwards sticks.
Your everyday Team Chat preferences live in Team Chat itself: Team Chat settings, the first item in the sidebar header's ⋯ menu, opens the settings window — message style, profile photos, how direct messages are sorted, which key sends a message, and the highlight-to-quote button — plus links to notification preferences, your workspace profile and (admins) workspace settings. Every control there is the same setting the main Settings pages show, so the two never disagree. In the main Settings, the Team Chat toggle sits in Settings → Lab, and Settings → Team Chat notifications carries the global default and the per-channel overrides — reached from the settings window's Notification preferences… link or a channel's bell menu. On a phone the same chat is reachable two ways: through Omniscio's own mobile web access while your desktop is running, or through a standalone phone web app that works with the desktop off. Through the mobile web access link, the Mission Control actions (Send to Mission Control, # item mentions) are hidden, because they only work in the desktop app. Small controls grow to a 44-pixel tap size on a phone, and the media gallery keeps its buttons visible on any screen that cannot hover, tablets and phones turned sideways included. Through the mobile web access link, guest invitations are not listed: the sidebar shows the note Workspace invitations are accepted on your computer. Open Omniscio on your desktop to see and accept them. instead, because viewing, accepting and declining an invitation are desktop-only. The standalone phone app is ready for translation: every word it shows comes from one English list inside the app, and it picks a translated list automatically from the phone's own language once one is added. Today it ships English only, because no translated list has been added yet, and a change only reaches phones after the phone app is uploaded again.
The assistant can point at every Team Chat control. Every button, link and clickable row in Team Chat's desktop window carries a named pointer (data-ui-anchor), declared beside its component in that component's own *.ui-anchors.ts file, so when you ask the assistant where something is it can find the control and draw a ring around it. A control that repeats (a reaction pill, a channel's Join button, a saved message) uses one name for every copy. A small wrapper that draws its own button takes its pointer through a uiAnchor prop. The coverage check keeps the whole Team Chat folder at zero unpointed controls: Team Chat has no rows left on its list of excused files, so new Team Chat code without pointers fails the build.
Enabling it
Open Settings → Lab and toggle Team Chat on (or set the AMC_SHOW_TEAM_CHAT=1 environment variable). It ships dark by default — the workspace must be set up before it does anything useful, so enabling it with no workspace simply shows the "not in a workspace yet" state — which carries a "Create a workspace" button (the same dialog the sidebar switcher's + opens), so a brand-new user with no admin to ask can start a self-serve workspace from that screen rather than reading a dead end.
The Lab toggle brings Team Chat's related tools with it; the environment variable does not. There are 17 of them — PM item mentions and the other Mission Control links, channel invitations, calls, the Slack and Discord imports, the Team Library, and the agent watchers. Each one reads the visibility of Team Chat above it, so with Settings → Lab they appear alongside it when their own switch is on, which is what the Lab help text promises. A developer launch with AMC_SHOW_TEAM_CHAT=1 reveals Team Chat alone and leaves every one of them looking as though Team Chat were still off, because the shared gate checks the parent with the add-on's own reveal flag rather than the parent's. The fix belongs in the shared gate (unreleasedFeatureVisibilityReason in src/shared/unreleased-features.ts): check the parent with the parent's own reveal. Until then, use the Lab toggle when the related tools are wanted.
Paid AI calls and who they are charged to
Sending, reading and syncing chat costs no AI money. Three Team Chat features do call a paid AI, and each one is recorded in the app's spend log (Stats → Spend) against what caused it:
| Paid call | What starts it | Charged to |
|---|---|---|
| Omni channel summary | An Omni session asks for one channel's summary | The session that asked, labeled Team Chat channel <id>, under the omni-team-chat-summary spend tag and its daily cap |
| Mission Control bot answer | A person types @mc in a private channel |
The feature and the signed-in account, labeled Team Chat channel <id>, under the pm-mc-chat-bot-inbound spend tag and its daily cap |
| Starter bot | A message matches a channel's starter bot | The session it starts, paid like any session; the spawn is linked to its channel |
Why a bot answer has no session. A chat message is not a session. Nothing that runs inside Omniscio asked for the answer, so there is no session to charge. The charge goes to the feature's own spend tag and the account that paid for it, and the channel label says which conversation it came from.
Tracing a starter-bot session back to its channel. Each spawn writes one row to team_chat_bot_session_links (spawn row id, channel, workspace, reply mode). Join it on sessions.origin_spawn_id to get the session and its normal cost. A spawn that waits for an approval in the inbox writes no row, because no session is started until someone approves it.
The labels carry only the channel id, never the channel name or any message text, and they stay on this computer. Contract: central-ai-spend-contract.md a-feature-becomes-drillable-in-one-line.
How it behaves
What you can do (v1)
Channels — create public channels (any teammate can read and post) or private channels (only invited members). When you create a private channel you pick who is in it and give each person a level: Owner, Member or Read-only, so a private channel can start with a co-owner instead of you promoting someone afterwards. A private channel shows a 🔒 lock beside its name (hover it for what it means), and it is invisible everywhere to people who are not in it — the sidebar, Browse channels, search and
@-mentions. A workspace owner or admin who is in the channel can switch it either way from its right-click menu (Make private (invite only) / Make public); going public asks first, because everyone in the workspace then sees the channel and everything already posted in it.The sidebar header — the top row is your photo (with its presence dot), the workspace switcher on a phone, then search, + and ⋯; the Chats | Sessions switch sits on its own full-width row just under them, so it never collides with those buttons on a narrow phone screen. Click your photo to set a status, pick your availability or edit your profile; there is no separate profile row under the header, on desktop or on a phone.
Sessions in Team Chat — the Sessions tab is the same session list every project has (needs you, running, archived, + New session, the right-click menu), with a running / waiting count on the tab. It lists Team Chat's own sessions and every session you started from a message, a thread or the Starter Bot, even though those run in the project you picked. Opening one keeps you in Team Chat; Chats takes you back to your conversations. Type
/session <task>in a conversation to start a session on that task: it is named after the conversation, links back to it, and posts nothing to the channel. Contract: D31 in team-chat-desktop-contract.md.One + for every new conversation — the sidebar header holds search, + and ⋯. The + opens one small menu — New session, then New channel (hidden for guests), New direct message and Connect by email — so there is a single place to start anything; no other menu repeats them. The ⋯ menu always shows and is grouped with dividers: Team Chat settings first, then people (add members, invite), then the workspace (workspace settings, custom emoji), then imports — a group with nothing in it for you is left out along with its divider.
A sidebar that only shows what earns a slot — Custom Emoji lives in the sidebar header's ⋯ menu, Browse channels and New channel live in the ⋯ menu beside the Channels heading (shown even before you have any channels), and the Saved row appears only once you have saved something. Details + invariants: team-chat-needs-you-and-dm-topics-contract.md.
Filter the sidebar — a Filter channels and DMs box opens at the top of the sidebar list when you pick Filter channels and DMs (the first item in the sidebar header's ⋯ menu) or press Alt+Shift+F; it stays out of sight the rest of the time (the 🔍 icon in the sidebar header searches messages). It stays open while it holds text and tucks away once it is empty and you move off it. Typing narrows Channels, Direct messages and Archived to names containing what you typed (capital letters and accents are ignored); the ✕ or Escape clears it, and ArrowDown moves into the first matching row. The same key puts the cursor in the box from anywhere in Team Chat (rebindable in Settings → Keyboard; the tooltip shows the current key). When nothing matches, the sidebar says No matches. The filter only hides rows — unread badges and the Needs you count never change — and it starts empty each time Team Chat opens. On a phone it works the same way: the box takes no row until you open it from ⋯.
Direct messages ⋯ menu — the Direct messages heading has its own ⋯ menu for how the list looks: Sort by recent activity (the default — conversations with something new for you first, then newest first) or Sort by name (A to Z), and Show profile photos. The sort choice is saved (setting
teamChatDmSort) and follows you across restarts. The heading and its menu stay visible even before you have any DMs.Needs you — one list of what is waiting on you — the Needs you row in the sidebar opens the list for the open workspace and shows its count; the inbox button on the workspace rail opens the same list across every workspace. It comes in three groups — Connection requests (accept or decline in place; they belong to no single workspace, so they appear in the all-workspaces list only), Replies owed (DMs where the other person spoke last and you have not answered, each previewing their last message) and Mentions & replies — one row per conversation or topic, so its count equals the rows you see. Mark done clears a reply owed; a new message brings it back. A middle click on a Replies owed row does the same as its Mark done, and on a Mentions & replies row the same as its Dismiss, so a row can be cleared from anywhere on it instead of hunting for the small button; connection requests are left alone, because accepting or declining asks first and cannot be undone. Answering inside a message's thread counts as answering it: reply in the thread under a question asked in the main conversation and that conversation stops saying it needs you (its notice and unread clear too). A reply in a different topic, or under an older message while a newer one of theirs is still waiting, does not. Replies owed start counting from when this view first ran on your device, so old conversations never flood it. The inbox button at the top of the workspace rail (Alt+Shift+A; on a phone, an All workspaces inbox entry in the workspace switcher) opens the same list, each row tagged with its workspace, and each workspace tile's badge is that workspace's share of it. The rail inbox button and the app's main-navigation Team Chat badge always show the same all-workspaces number, the sidebar row shows the open workspace's share of it, and all three wear the same soft amber pill as the app's main Inbox count. Alt+Shift+I opens the open workspace's list from the keyboard. Each mention row has a Dismiss button, and Mark all done at the top clears every reply owed and mention in the list at once (connection requests are never answered for you); a toast offers Undo. Dismissing a mention also marks that conversation read, so the unread badge and the workspace tile's count cannot outlive the row that explained them; a message that arrived after the dismissed notice keeps the badge, because it is still waiting on you. While the list is still loading it says so instead of claiming you are all caught up. Archiving a DM's Team Chat notice in the Inbox marks that conversation done — whether you, a command-line action or your agent put it away — and restoring the notice takes the mark back; an archive the app makes on its own never does. A connection request you snoozed in the Inbox stays out of every count until the snooze ends.
History — get back to what you already read — reading a conversation, Dismiss and Mark done all take an item out of Needs you, so the Needs you header has a History button that drops down a Recently read list of what left it: one row per conversation (its newest message, reply or reaction), showing the real message text, most recently read first, with a workspace label in the all-workspaces list. Click a row to reopen it: a thread reply opens the exact thread it announced, anything else opens the conversation (a reply notice saved before notices remembered their thread also opens the conversation, never a guessed thread); the dropdown closes, and Esc or a click outside closes it too. A conversation that needs you again moves back to Needs you and leaves History, cancelled reminders never appear, and History rows are never counted — every Needs you number is unchanged. It is the whole record of what left Needs you on this computer, whatever has happened to the conversation since: one you archived, left, or that is simply not in the sidebar right now still appears here, named as it was when it arrived. The list is warmed while Needs you is on screen, so opening it shows your history straight away instead of a spinner. A load that fails says so, and only an empty history says Nothing read yet. Works the same in the all-workspaces inbox and on a phone. Details + invariant: team-chat-needs-you-and-dm-topics-contract.md
needs-you-history-keeps-what-you-read.Previous / Next — step through what needs you like the inbox — only on a trip you started from Needs you (opening one of its rows, or its History): while you are in the conversation that trip took you to, its Topics list or one of its threads and anything is waiting on you, a thin bar under the header reads ‹ Previous · Sam Lee 2 of 5 · Next › — the name is whoever the Next arrow opens, the count is plain grey text, and on the last item only the count shows. The count never claims more than really waits: once some in the list are answered it adds how many still wait (2 of 5 · 3 waiting), and once all are answered it reads All caught up with no name. Next (Alt+Shift+J) and Previous (Alt+Shift+K) open the item after / before the one you are on — both rebindable in Settings → Keyboard, shown in the button tooltips, and both work straight from the message box; each is greyed out at its end of the list. The order is fixed so it never reshuffles as you move: this workspace first, each person's waiting threads together (including older topics with an amber dot), people ordered by their newest waiting thread, newest first, then other workspaces — switching workspace for you when it has to. A thread you open stays in the list, so Previous takes you back to it and the count does not shrink just because you looked; something new that starts waiting is added at the end. On a screen outside the list the bar reads 3 waiting and Next opens the first one. It disappears when nothing is waiting, and the trip ends the moment you open a conversation any other way — clicking a person in the sidebar or the Direct messages list, a Message button, search or a notification never shows the bar, and the next trip from Needs you starts a fresh count. It never moves you on its own: once you have answered the thread you are on — your message is now the newest, or you chose Mark done — the Next arrow lights in the accent colour (the line itself stays grey), and you press Next when you are ready. Simply opening a thread does not count as done. Connection requests stay in the Needs you list. Details + invariants: team-chat-needs-you-and-dm-topics-contract.md
up-next-names-what-needs-you-next·up-next-never-moves-you-on-its-own.Browse & join channels — a "Browse channels" entry in the Channels ⋯ menu opens a searchable directory of every PUBLIC channel, each with a member count. Open any channel straight from the directory (public channels are readable without joining, like a Slack preview), or Join / Leave to add/remove yourself from its member list. Every public channel is also listed in the sidebar for every workspace member, joined or not (guests see only their own channels): your joined channels come first, the rest follow, and only joined channels raise unread dots, badges or notifications. Leaving a public channel stops its notifications but keeps the row. To take a public channel you are not in out of your own sidebar, right-click it and choose Hide channel: only your list changes, on every computer you sign in on. Hidden channels stay in Browse channels, marked Hidden, where Show in sidebar brings the row back; joining a channel, or being added to it, shows it again too. The Team Chat phone app leaves out the same hidden channels; hiding and showing them is done on a computer.
A ⋯ on every row — every channel, DM and archived row has its own ⋯ button that opens the same menu as a right-click; the one row without it is a cross-company conversation. On a computer it appears when you point at the row; on a touch screen it is always showing.
An empty workspace still has a way forward — a workspace with no channels and no DMs says No channels yet with New channel and Browse channels buttons (a guest gets New direct message instead), and the Direct messages heading stays below it.
Synced stars — starring a channel saves it to your account, so your starred channels follow you to your other devices. The first time a computer opens a workspace after this change, the stars it kept on its own are added in once (never replacing what is already there).
Sent requests and Cancel — connection requests you sent sit under a Sent requests heading near the bottom of the sidebar, each with Waiting… and a Cancel button. Cancel takes the request back: the other person no longer sees it and you can ask the same address again straight away. A request they already accepted stays connected. Guest invitations sit above it, and Archived comes last.
Manage channels (right-click) — right-click any channel in the sidebar to open a context menu with Edit channel (rename or change the topic), Leave channel (public channels only), and Delete channel (permanently removes the channel and all its messages — confirms first). Edit and delete are available to the channel's creator or a workspace admin; leave is available to any member of a public channel, except that a channel's only manager is refused ("Cannot leave as the last owner. Promote another member first.") until they promote someone else, both in the app and through the CLI leave route. Uses the same
MenuShellcontext-menu scaffold as the rest of the app; the edit dialog pre-fills the current name and topic. Invisible, direction-flipping and control characters are stripped from a channel name when you save it, and the server refuses any create or rename that still carries one, so no channel can be named to look like another (invariantchannel-names-carry-no-hidden-characters).Workspace settings — for an owner or admin of a workspace you created, the window has two tabs: General (the workspace name with a Rename button, and its icon) and Members (the roster, roles and removal — the member list lives only there). Members of other workspaces see just the member list. An Add members button sits above the list: an owner or admin clicks it to close Settings and open the invite flow; everyone else sees it grayed out, with the reason when they hover or tab to it.
Channel members — the member count beside a channel's name (a people icon and a number) is the one way to see who is in it: it opens the channel's members panel when channel invitations are on, and the workspace member directory otherwise. A direct message keeps its Members button instead. The panel lists the members as it opens — the names come from the channel and the workspace roster it already holds — and only each member's role waits on a read, showing a short placeholder bar until it arrives.
Thread pane — a thread or topic has one way out: the Back arrow at the left of its header (with a tooltip). An untitled topic is headed Topic; a reply thread is headed Thread.
New direct message window — each teammate shows their photo (or initials) with an online dot. When nobody else is in the workspace yet, the window offers Connect by email to reach someone outside it, and admins see Don't see someone? Invite people.
Smaller wording fixes — your profile popover's row reads Edit my profile in this workspace (the same words as the workspace right-click menu), the Scheduled row stays visible while its list is open, and menu labels use sentence case (Edit channel, Archive channel, Leave channel, Custom emoji, Needs reply).
Make one (Emoji Maker) — when the Emoji Maker plugin is installed and on, a Make one button sits beside Add Emoji in the Custom emoji dialog and in the emoji picker's Custom section; it opens the plugin's panel to design an emoji. The plugin hands its picture back to the same add-emoji dialog with one row already filled in, and nothing is added until you press Upload. The button never shows on a phone. Details + invariants: team-chat-custom-emoji-contract.md.
Channel name and topic limits — Edit channel refuses a name over 200 characters or a topic over 2,000, with a message naming the limit, before anything is saved.
When a new channel is refused, the dialog says which limit you hit — New channel shows the server's own reason instead of a connection error: the daily caps (per person, per workspace, and the all-workspaces activity cap) each say so and that they reset at 00:00 UTC; the plan cap says to upgrade; and the 200-channel workspace ceiling says that archived and private channels count toward it and that a workspace admin can delete unused ones to make room. A refusal that arrives without a reason from Omniscio's own server reads as the app's general "Too many requests right now… wait a moment and try again" message. Every such refusal is also written to
main.logwith the server's reason, so support can tell which limit it was.Edit channel keeps up with the live channel — the dialog reads the channel from the same list the sidebar renders, so a rename, topic or visibility change another admin makes while it is open appears in the fields you have not touched. If you did change that same field, saving asks first — "Someone else changed this channel" — instead of quietly replacing their change, and a channel that was deleted (or that you were removed from) while the dialog was open says so and refuses to save.
Channel archiving (admin) — a workspace owner/admin can archive a channel (right-click → Archive), which makes it read-only (the composer is disabled; an archive banner is shown) and hides it from the default sidebar and channel browser. Archived channels appear in a collapsible "Archived" section at the bottom of the sidebar. Unread badges are suppressed on archived channels. An admin can unarchive to restore normal operation. The archive fields (
archivedAt,archivedBy) are schema extensions (.merge()), so the frozen Phase-0 schema is untouched. A CLI endpoint is atPUT /team-chat/channels/:channelId/archive. Invariants A1-A5: team-chat-workspace-admin-controls-contract.md.Direct messages — start a 1:1 or small-group DM. The DM is addressed by a deterministic id derived from the participants, so "message Alice" always resolves to one conversation — no duplicate threads.
Direct Messages is one list of people — no People / Threads switch and no "Needs Reply" block in the sidebar. Each row is a person (or a cross-company connection); clicking a person opens the newest of their threads that needs you, otherwise Main when it has messages you have not read, otherwise Main when that is their only thread, otherwise their list of threads — so a row with an unread count always opens on the unread messages. The Message button on a profile card or in the member list follows the same rule but opens Main instead of the list when nothing is waiting. Conversations where the other person spoke last sit at the top, newest first; once you answer one, it drops below them. Every number in Team Chat is one count. A row's number is how many of its threads and mentions are in Needs you (with Needs you in its tooltip and screen-reader text), each workspace tile adds up its rows, and the rail inbox adds up every tile plus connection requests — so the three always agree. The server's unread count never draws a number: once you have opened a conversation (with clear-on-open on), marked it read or marked it done, it shows nothing anywhere until something new arrives. A cross-company conversation carries no number, because nothing can mark it read; it still sorts to the top while the other person has the last word. Both read one list, so a reply Needs you shows is never missing from the row, even when the message was already read or sits in a topic. That covers any of their topics, not only Main and the newest side chat: each one-to-one conversation keeps, on the server, the last author and time of its newest 50 topics (
threadReplyTails), so an older topic they spoke last in still lights their row and appears once in Needs you, even after a newer topic you already answered. A group DM counts exactly the same way — a conversation with three or more people gets the same needs-reply badge, Needs you row and topic tails, and the person its row is named after is whoever spoke last. The oldteamChatDmInboxViewoption is retired: it is still stored but nothing reads it.Send & receive in real time — messages appear instantly for everyone in the channel. Your own message shows immediately (optimistic), and if you go briefly offline — or a send has to retry after a network hiccup — it lands without ever double-posting: the client-minted id is the message's own id, so a resend updates the same message rather than creating a second copy (identical on desktop and phone).
Conversations open already up to date — opening any conversation — from Needs You, the sidebar or the Inbox — shows its newest messages straight away, with no catch-up a moment later. In the background the app keeps your recently viewed conversations and everything in Needs You current, fetching only the messages you don't have (same on desktop and phone).
@mentions — type
@to open a teammate typeahead (arrow keys + Enter/Tab to pick); selecting inserts the mention, which renders inline as a highlighted "@Name" chip and is parsed into the message'smentions[]from the inserted token so notifications highlight you.The standalone phone app holds up on its own — a message that cannot be shown now costs only the conversation pane: the sidebar and the composer keep working, and the recovery clears the open conversation before reloading instead of reopening the same crash. Thread reply counts and an opened thread come from the thread itself, not from the messages the phone happens to have loaded, so a count is the real count and a thread lists every reply. Messages loaded with Load earlier follow edits and deletions made by other people, and a message deleted elsewhere drops out of the history. The 📌 pinned count and the pinned list read one list, so the badge can never disagree with what it opens. Same on desktop where the two share the rules.
Phone alerts through the desktop phone link — when you read Team Chat on your phone through Omniscio's mobile web access (the link to your running desktop app), a direct message or an @-mention that would pop a desktop notification also rings your phone, and tapping it opens that conversation — whether the app was closed or already open on the phone. Ordinary chatter in an all-messages channel stays a desktop notification only. Your Team Chat notification settings decide it: a muted channel or DM stays silent, a mentions-only channel rings only on a mention, a thread reply rings only in a DM or when it mentions you, and with content-free notifications on (the default) the phone shows just Team Chat and New message, never the sender or the text. Quiet hours, Silence and Presentation Mode keep the phone quiet too, and one message rings once even when the desktop and the phone both see it. The standalone Team Chat phone app keeps its own alerts, unchanged. See Mobile and web push notifications.
Needs-reply clears when you look, and never stays cleared forever (default on) — toggle
teamChatDmAutoClearNeedsReply(Settings → Notifications → Team Chat) and a DM stops showing as needing a reply as soon as you actually look at it: the conversation is open, the app window is in the foreground, and Team Chat is the panel on screen — not merely a focused window while you work elsewhere in the app — the rule that already marks a channel read. The conversation stays in your DM list — glancing at someone never makes them disappear from the sidebar — and a new message from that person brings the tag straight back, so a look never buries a conversation. A thread row clears only once you open that thread (an unopened reply shows as an "N replies" line, so it has not been looked at). Turn the option off to go back to confirming each DM by hand with Mark as done. The inbox preview embedded in a notice never auto-clears, since you never chose to open it.Mark as done hides the conversation; looking at it does not —
markDmThreadDone(the hover ✓, the right-click item, Mark done in Needs you, and reacting to a message) records a done acknowledgement, which also drops the row from Needs you. Marking a person's row done marks every topic of theirs that is waiting on you done too, and Mark as not done on that row brings them all back. The automatic clear records a seen acknowledgement instead, which only stops the "Needs reply" nagging. Both record which message they covered, so either one expires the moment a newer message arrives — an old "done" mark can no longer bury a person's later message. Nothing else removes a mark: it does not wear off with time and there is no limit on how many you keep, so Mark done on a conversation that has been quiet for months stays done until they write again or you mark it not done. Done marks are kept on this computer: the read position syncs to your other devices, but the done list does not, so the same conversation can still wait on you on another computer — the action's tooltip says so, since nothing else on that machine can explain it.Reactions — react to any message with an emoji: an "Add reaction" button in each message's hover toolbar (desktop) — or the quick-reaction row of the mobile long-press action sheet — opens the emoji picker (it opens toward the message from the button and closes as soon as you pick, so one click = one reaction), and your reactions appear as pills (the emoji + a count) under the message — shown only once a message actually has reactions — live for everyone in the channel. Your own reactions are highlighted; click any pill to react with that same emoji (Discord-style, in channels, DMs, group DMs and threads, in every workspace), and click it again to take yours back. Right-click a pill (or long-press it on a phone) to see the list of who reacted; on desktop, hovering a pill also shows the names. You can react to your own messages too — a pill under one behaves exactly like any other. See Reacting to your own messages. Your reaction shows instantly and reconciles with the server (a failed action undoes itself); if it can't be applied — you're offline, or chat is still reconnecting — you get a brief notice instead of a silent no-op, so a click is never lost without explanation. Reaction-author names in the pill tooltip light up once the workspace member directory lands; until then a pill shows the count. In a DM, adding a reaction to a message someone else wrote also clears the "needs reply" indicator — the same effect as the "Mark as done" checkmark — and, like every other acknowledgement, it expires if that person sends something newer.
Pins — pin any message to the channel: a pin button on each message toggles it pinned/unpinned, and pinned messages collect in a "Pinned" bar at the top of the channel (a count that expands to the list, newest first). Any channel member can pin or unpin — the bar and each message's pin state are live for everyone in the channel. Your pin shows instantly and reconciles with the server. A pinned message that has scrolled out of the loaded history is fetched on its own: its row holds a one-line placeholder while that read is in flight, then shows the author and the text — you no longer have to page "Load older messages" to read a pin.
Threads — reply to a specific message to start a thread. Each message has a "Reply" action in its hover toolbar (desktop) or the mobile long-press action sheet, and once it has replies it shows an "N replies" indicator that reads Slack-style — a small facepile of the repliers' avatars sits beside the count, with the count in the accent colour so it stands out as a clickable link (making a busy thread easy to spot). Click either to open the thread pane (it fills the message pane at every width, phone and desktop alike — a thread is a view of its own, not a strip beside the conversation) — it shows the original message, its replies in order, and a composer for your own. The pane opens on its newest reply and stays with the conversation: sending a reply always scrolls it into full view, wherever you were reading, and someone else's reply scrolls in only while you are already at the bottom — reading older replies is never interrupted. The conversation's floating up/down arrows hide while the pane is open, so they never cover it. A reply appears instantly (optimistic) and is live for everyone; if a reply can't be sent it flips to a failed row with Retry / Delete (the same safety net as the main composer) instead of sticking on "Sending…" forever; the main channel list stays clean (it shows only the original messages, never the replies). Leave the pane three ways: the back arrow at the left of its header, the X at the right, or Ctrl/Cmd+W. All three land in the same place — the topic list when you opened the thread from it, otherwise the conversation behind it. The back arrow matters once the pane covers the conversation — it is the only control at the leading edge, where a way out is looked for. And typing an ordinary letter (like
eorw) in the reply box never closes it.Side replies inside side replies — every reply inside a thread has its own Reply, as many levels deep as you like, and a reply that has its own replies shows its own "N replies" count. In a private channel or a DM, Reply opens that reply's own thread right there; the back arrow and Escape step up one level at a time (the X and Ctrl/Cmd+W still close the whole pane). In a public channel, Reply on a message inside a thread is private by default: a Private side reply picker opens with the message's author and anyone it @mentions already checked (never you, never guests, never @everyone), you add or remove people, and Start private reply opens a private group chat with exactly those people on a fresh topic, with a link back to the original message already typed in. To answer where the whole channel can see it instead, pick Reply in thread (everyone) from the message's menu. A message's reply count always opens its existing replies. Guests reply in place, since they cannot use private chats. Older copies of the app show only first-level replies until they update.
Topics — separate named conversations, browsed like an inbox — a chat is organized into topics. Clicking a person opens the newest thread that needs you, else Main when it has unread messages, else Main if that is their only thread, else their Topics list; a channel, or a notification or link to a specific message, still opens the conversation itself. Every person-DM screen has a back arrow next to the person's name (the header reads person › Main or person › topic), and each press climbs one level: a thread goes to that person's Topics list, and the list goes to everyone you DM with — the chat list on a phone, or on desktop a full-pane Direct messages list (same order as the sidebar, with a Needs you tag on anyone waiting on you). Alt+← does what the visible back arrow does (never while you are typing). Arriving from the inbox: when an Omniscio inbox notice, or a chat link followed while the app is already running, opens a conversation for you from outside Team Chat, the first press of that arrow returns you to where you came from — the Inbox, or the project the link arrived from — and the arrow says where it goes ("Back to the inbox"). That return belongs to the one jump that created it: pick any conversation yourself, from the sidebar, the Direct messages list or a person's row, and it is spent and the usual ladder applies. A link that launched the app creates none, and a phone keeps its own ladder. In a DM the arrow replaces the Topics button; channels keep it. The list always starts with a Main row — the ongoing conversation — then the topics you pinned (the pin button on each row; pins are yours alone, follow you to your other devices, and hold up to 100), then the rest. From inside a conversation, a Topics button in the header steps up to that list of the conversation's topics, read like email: one row per topic with its name, who spoke in it last, what they said, and when. Click a row to open that topic — and closing it returns you to the list, not to the main thread. Starting a topic is one click from anywhere: a + New topic button sits in every conversation header — beside Topics in a channel, and alone beside the name in a person DM, where the back arrow replaces Topics (it shows even before the conversation has any topics), and opens an empty topic straight away — no need to open the list first. A topic you started but have not sent into yet stays put: open another chat (or a connection, or another workspace) while you are in it and the next time you open that conversation you land back in the same topic with your half-typed message still in the box. Stepping back out of it with the back arrow first lets it go. It is remembered until the app restarts. Alt+Shift+N does the same from the keyboard — rebindable in Settings → Keyboard, and the button's tooltip shows the current key — and Ctrl+T starts one too, from anywhere in Team Chat, including while the message box has focus. Ctrl+T is Team Chat's own key: outside a chat it still opens a new session. A topic opens instantly under a name taken from its first message — no typing a title — and a moment later AI renames it to something clearer, on the same engine and spend lane your session titles already use. It only ever names a message you wrote: a topic started by replying to someone else's message keeps its first-line name, because nobody's words go to your AI but your own. Whatever the name, anyone in the conversation can rename it, and a name you type is never overwritten by the AI naming run already in flight. When that first message @mentions someone, the topic name, its preview and the Needs you row all show the person's name — never their raw user id — and a rename keeps the mention linked to them. Replying to a message makes a topic too, starting with the message you replied to, and it appears in the same list — so a side conversation from weeks ago is findable by name instead of only from the message it hung off. Works in DMs and channels alike, and on a phone (where the button drops its label so it cannot crowd the name beside it). An agent can start one too: sending a message through the command line with
newTopicset makes it the first message of a new topic, named the same way, and a retried send lands in the same topic rather than a second one (team-chat CLI). Topic messages stay out of the main conversation, and the chat list still sorts on real activity. Clearing out a dead topic: right-click a topic row (or press its ⋯ More actions button) and choose Hide topic to take it off your own list on every device you use. Nobody else is affected, and the topic comes back by itself as soon as someone writes in it. The person who started a topic, or a workspace admin, also sees Delete topic. After a confirmation ("This deletes this topic for everyone in this conversation."), it removes the topic from the list for everyone, while its messages stay stored. The Connection chat row offers Hide only, because it is your connection with that person rather than a topic. Hidden topics follow you between devices only once the Team Chat database rules that allow them are deployed; until then, Hide shows a plain error and nothing changes. Putting a finished topic away: the row menu also offers Archive topic and Snooze…, the same two put-away actions a session row offers. E archives the topic you are reading (rebindable in Settings → Keyboard, shown in the menu item's tooltip, and it never fires while you are typing), and middle-clicking a topic row archives it, with an Undo toast. Archive takes the topic off your own live list on every device and keeps it on an Archived (N) shelf at the bottom of the topics list until you take it out — it comes back only when you choose Unarchive topic, never by itself, because a topic that needs your reply still appears in Needs you. Snooze… opens the app's snooze picker with the same choices every other snooze has (later today, tomorrow, next week, a time you type), hides the topic until then, and brings it back on its own; until it is due it waits on a Snoozed (N) shelf, each row naming when it returns, with Unsnooze. Both shelves reach past the topics listener's newest-50 window: the same backfill that already rescues a pinned topic fetches a snoozed or archived one too, so a topic put away months ago is still there. Both are yours alone, are stored on your own saved document beside your pins, and follow you to your other devices, with the same caveat Hide carries: until the database rules that allow them are deployed, the write shows a plain error and nothing changes. Details + invariants: team-chat-topic-threads-contract.md.Rich composer — a formatting toolbar inserts markdown (bold, italic, inline code, fenced code block, block-quote, bulleted/numbered list — plus Ctrl/Cmd+B/I/E) and an emoji picker inserts emoji at the cursor. Pasting rich content — a formatted outline, a Google-Docs selection — drops in as clean Markdown (bold, lists, headings, and links preserved) at the cursor, honoring your Paste rich text as Markdown setting (hold Shift, or Ctrl/Cmd+Shift+V, to paste plain); a plain-text or file paste is left untouched (files still stage as attachments) — the same behavior the main session composer has. Send key is your choice (Settings → Sessions → "Team Chat submit key"): Enter to send (default — Shift+Enter for a newline) or Ctrl/Cmd+Enter to send (Enter for a newline); Ctrl/Cmd+Enter always sends in either mode. On a phone or tablet, pressing Enter to send also closes the on-screen keyboard (the standalone Team Chat web app does the same); on a computer the box stays focused for your next message. List auto-continuation: pressing the newline key (Shift+Enter in Enter-to-send mode, plain Enter in Ctrl+Enter mode) on a bulleted (
-) or numbered (1.) list line auto-inserts the next marker; on an empty marker it strips it to exit the list. The box grows as you type, and when the cap is not enough, a drag handle above the box lets you pull it as tall as you want (desktop only — drag it back down to the resting size and the box goes back to growing with your text; the handle is also reachable with Tab, and Arrow Up / Arrow Down grow or shrink the box in small steps with the same limits as dragging), and the composer rests as a single row — the formatting toolbar is revealed on focus on desktop (so an idle composer never parks buttons above it) and stays behind a small toggle on a phone (where the soft keyboard already crowds the screen); every button and keyboard shortcut still works whether the strip is showing or not. The box is the app's shared editor (Settings → Sessions → "Team Chat uses the new shared editor", on by default): a real document with real lists, so Tab nests an item and Shift+Tab un-nests it, the newline gesture continues the list at the same depth (Shift+Enter in Enter-to-send mode, plain Enter in Ctrl+Enter mode), and on an empty bullet steps back one level (leaving the list only at the outermost level) instead of dropping you out of it — and an ordered list counts up the way it will be sent, with no letter markers to rewrite on the way out. Your send key is unchanged: in Enter-to-send mode Enter still sends from anywhere (Shift+Enter edits the list), and in Ctrl/Cmd+Enter mode Enter edits the list and never sends. Turning that setting off brings back the previous message box, live, no restart. A quote shows itself in the box — a line that starts with>becomes a real quote, drawn with its own>marker and bar at every level, so a quote you paste reads as a quote where it landed and a>you type on a quoted line (which nests it one level deeper) is visible the moment you type it. Leaving a quote: the newline key on a quoted line starts another quoted line, and pressed again on the empty quoted line it takes you out of the quote — so the reply you type under a quote is never pulled into it, in every editor style. Reply focus matches a session's — opening a channel, DM, or thread drops your cursor in the box from the start (when you keep the default auto-focus on), and pressing R focuses it when auto-focus is off — the same behavior you get on a session (skipped on mobile so the keyboard doesn't pop up).Voice dictation & FlowVoice — dictation is FlowVoice, and only FlowVoice: hold its hotkey (or click the FlowVoice mic in the composer) and the transcript is typed at the OS cursor, landing in the Team Chat box when that is focused. The older in-app dictation (wake word, mic button, Alt+V, push-to-talk, and the shared Dictation delivery setting) has been retired. FlowVoice auto-submit routes to Team Chat when it is the active panel, sending the pasted transcript through the real send path. A FlowVoice mic button (
FlowVoiceMicButton) is rendered in the composer's action bar (between the emoji picker and the preview toggle) so users can start/stop dictation with a click without needing the hotkey; it self-hides when FlowVoice is disabled or on mobile (desktop-only). No new settings are needed.Formatted messages — what you format actually renders: bold, italic,
inline code, fenced (syntax-highlighted) code blocks with a hover-revealed copy button (top-right overlay; copies clean text viatextContent), quotes, bulleted/numbered lists, and clickable links all display formatted in the message — through the same renderer the rest of the app uses for messages (ChatProseBlockinherits the sharedmarkdownComponentspipeline, including theCodeBlock+CopyButtonoverlay). @mentions still show as highlighted "@Name" chips, and an@nametyped inside code stays literal.Message layout — pick your style — choose how the message list looks in Settings → Appearance → "Team Chat message style": Bubbles (the default) shows rounded 3D bubbles exactly like the regular Omniscio chat — your own messages sit on the right in an accent bubble (just the time — your name/avatar dropped, since it's you; an imported message of yours still shows its Imported tag), everyone else's on the left with their name + avatar, each bubble only as wide as it needs and never more than three-quarters of the chat pane (so even a long message stops short of the edge on a docked pane). The small (24px) picture, name, Agent label and time sit on the first line inside the bubble, so there is no empty column beside or under a message. That first line is capped at the bubble's width, so a long name or session title wraps onto at most two lines and then shows "…" instead of widening the line — the picture, name, Agent label and time all stay on screen even on a phone, where an over-long title would otherwise run off the message's edge on the side its bubble is aligned to. Bubble colours are four shared tones, the same in Team Chat and the Help Desk (your regular sessions keep their own grey agent bubble with a green left stripe): your messages are blue, other people's are grey, your agents have a faint blue wash with a thin green outline, and other people's agents are grey with the same green outline — and an agent's picture carries a small robot badge; the phone app paints these same four, so your agent and a teammate's agent never come out as one bubble there either; Clean is one calm borderless single stream; Compact is dense. Messages are grouped by sender — a run of messages from one person shows the avatar/name/time once, not on every line. The look stays calm: each message's actions — react, reply, pin, save, copy, and edit/delete your own — live in one small toolbar that appears on hover on desktop — it rests as a single ⋯ button and unfolds into react, reply, start-a-session and edit when you hover or focus it, and it follows a long message: while any part of the message is on screen its actions stay pinned near the top of the message area instead of scrolling away with the message's first line; on a phone the toolbar is gone, and a small ⋯ sits beside each message — tap it (or long-press the message) for a bottom action sheet with the same actions. Reactions + the "N replies" indicator show only when they exist. The bubble's 3D depth follows your Chat Depth setting and its colour follows your theme (nothing new to configure). The same layout carries into the thread/reply view.
Real names & avatars — every message from a teammate shows the author's real display name and photo (with an automatic initials fallback), resolved live from the workspace member list — never a raw account id. (Your own messages drop the name + avatar.)
Agent messages are clearly labeled — everywhere, and impossible to hide — when a message is posted by a teammate's agent (an AI automation acting on their behalf — e.g. via the command-line relay route
POST /team-chat/channels/:channelId/messages, or a channel's Session Starter Bot acknowledgment) rather than the person typing in the app, it is unmistakably marked as "<Name>'s agent" plus a 🤖 agent marker — on desktop, the robot badge on the author's picture, or a small labeled "Agent" badge only where the row shows no picture (the compact message style, or the inbox conversation view); never both, since the badge beside a robot-marked picture is redundant — so a teammate's agent, not the teammate, always wrote it. Because an agent posts under the person's own signed-in identity (authorUid), whose agent it is follows automatically — and the label appears on every surface the message is read: the desktop message row (including the viewer's OWN agent's messages, and inside the inbox conversation view — the two places a normal own/embedded message hides the name), the mobile app, and the notification + inbox preview line ("Alice's agent: …", for channels, 1:1 DMs, and cross-org connection DMs). An agent reply also breaks the sender grouping, so it can never tuck in silently under the person's own messages as an unlabeled continuation (grouping by author alone couldn't tell an agent reply from the person's own — they share one identity). An un-labeled agent message is basically impossible: all three agent-only send relays — a channel post, a 1:1 DM, AND a cross-org connection DM — stamp the agent-authorship flag unconditionally, and they stamp it server-side, inside the relay that writes the message (they are the agent-only send paths; humans write directly to Firestore and never touch them). A client write cannot carry the flag at all — the Firestore rules refuse any ID-token write that sets it — so no agent send path can forget the stamp or forge it. An agent message also carries a distinct agent bubble — a thin green outline (thestatus-agenttone) a person's bubble never has, plus a robot badge on its picture — so an AI reply is set apart at a glance as well as by the label, on desktop and the mobile web app. A normal, human-typed message is unchanged. Backed by the additiveauthoredByAgentmessage field; the "<Name>'s agent" labeling is one shared helper across all surfaces. Invariants: D22 in team-chat-desktop-contract.md ·an-ai-agent-reply-is-identified-on-mobilein team-chat-pwa-contract.md · field in team-chat-schema-foundation-contract.md. Agent traffic also draws on its own allowance, separate from your own — both the daily limit and the per-minute rate that decide whether a message is notified about. A busy fleet of agents can never use up how many messages you personally can send, or be notified about, in a day or in a minute, and every "you have reached today's limit" reply says which allowance ran out. Machine traffic counts as machine traffic wherever it comes from — the posts a channel's webhook, alert-routing or Session-Bot ingest generates are written as you but draw on the same separate machine allowance, so they cannot crowd out your own. Invariant:agent-traffic-has-its-own-daily-counterin team-chat-messaging-relay-contract.md; the fan-out gates are in team-chat-notifications-contract.md.The Agent badge is a link back to the session that sent the message — the green "Agent" badge does more than label: clicking it opens the Omniscio session that posted that message, so reading an agent's update in chat is one click from talking to the agent itself. It navigates exactly like every other "go to session" link (an archived session, or one in a project you aren't currently in, opens the same way). It works because the machine that sent the message already recorded which of its own sessions authored it, in a local ledger keyed by message id — and that fact is deliberately never written onto the shared message, because a session id only resolves on the computer that owns the session. So a teammate reading your agent's message gets the same badge with no link, rather than a dead one. The badge looks and reads exactly as it always has whenever no sending session is known: another machine's agent, a bot, a scheduled send, or a message older than the ledger. On a message where it IS a link, the badge makes that obvious — it takes a pointer cursor and the standard focus ring, so it never reads as the static label beside it — and hovering it names where it goes, showing the sending session's own title ("Open session Fix the login redirect") rather than a generic hint, so you know which session you're about to land on before you click. The name beside the badge goes to the same place: on your own agent's message it shows that session's title, so clicking the name opens the session too (it never opens your profile card). A teammate's agent message, where no session is known here, keeps its name opening that person's profile. Invariant:
an-agent-message-points-back-at-its-sessionin team-chat-desktop-agent-invariants-contract.md.A plugin's post names the plugin — when a plugin posts in Team Chat or sends a direct message under your name, the message reads "Alice · sent by Notes" (your name, then the plugin's name) instead of "Alice's agent", and it keeps the same robot badge or "Agent" label, so everyone can tell which plugin wrote it. The plugin name is shown as plain text, up to 60 characters, and only Omniscio's own send service can attach it, never a message someone types in the app. On your own computer a plugin's post keeps this wording too, rather than showing a session's title the way your own agent's posts do. This shows once the latest Team Chat server update is live; until then, a plugin's post still reads "Alice's agent". Invariant:
a-plugin-post-names-its-pluginin team-chat-desktop-agent-invariants-contract.md.Agents can be blocked from Team Chat writes entirely (owner opt-in). Attribution above makes an agent send visible; a separate second layer can refuse it. The
allowAgentTokenTeamChatsetting (Settings → CLI Control, off by default) governs whether an agent's scoped$AMC_CLI_TOKENmay write to Team Chat at all — send a DM or channel message, react, edit, delete, pin, schedule a DM, invite a guest. Off ⇒ every gated write over the CLI from a scoped agent token is refused (403), so a compromised MCP server or npm dependency that reads an agent's env can't message your teammates as you — mirroringallowAgentTokenSpawn. The same switch covers an agent running on a cloud box (its own cloud token): a cloud session can do in Team Chat exactly what a local agent can, no more. <!-- team-chat-agent-writes:open -->Two writes are always open to an agent's own token and need no switch, because they post nothing and read nothing: unsubscribing (DELETE /team-chat/subscriptions/:id) and stopping a reply watch (DELETE /team-chat/reply-watches/:channelId).<!-- /team-chat-agent-writes:open --> With it on, an agent's own token can also <!-- team-chat-agent-writes:switch -->subscribe to a person or a thread (POST /team-chat/subscriptions), react (POST /team-chat/channels/:channelId/messages/:messageId/reactions), schedule a DM (POST /team-chat/dm/schedule), cancel a scheduled DM its own session created (DELETE /team-chat/dm/scheduled/:id), post a channel alert (POST /team-chat/channels/:channelId/alerts), create a channel (POST /team-chat/channels), send a channel message (POST /team-chat/channels/:channelId/messages), send a DM (POST /team-chat/dm), and — cross-org — request, respond to, disconnect and withdraw a connection (POST /team-chat/connections,POST /team-chat/connections/:id/respond,POST /team-chat/connections/:id/disconnect,POST /team-chat/connection-requests/:id/withdraw), send a connection DM (POST /team-chat/connections/:id/messages), set connection preferences (PUT /team-chat/connection-prefs) and toggle a channel's AI switches (PUT /team-chat/channels/:channelId/ai).<!-- /team-chat-agent-writes:switch --> Editing or deleting a message, pinning, saving and marking read stay closed to agent tokens. The global CLI token is always allowed; for an agent's own token the same switch gates Team Chat reads as well as writes. This is the layer that covers the react/edit/delete/pin vectors a label can't mark. Invariant: D22 (e) in team-chat-desktop-contract.md; gating in cli-server-gating.md.Edit & delete your own messages — on desktop, right-click any message for the full action menu (Reply, Pin, Save, Copy text, Copy link); your OWN messages also show Edit (inline, shows an "(edited)" marker) and Delete (confirms, then offers an Undo that re-posts it) below a separator. On a phone, tap the ⋯ beside any message (or long-press it) → the action sheet with the same actions — in channels, DMs, threads, and DMs with someone from another organization alike. A long sender name shortens with "…" rather than letting the timestamp spill over the ⋯, so the button is always visible. You can only edit/delete your own messages.
Copy message link — every message has a share icon (desktop hover toolbar) or a "Copy link" row (mobile long-press action sheet) that copies an
omniscio://chat/…deep link to the clipboard. Paste it anywhere — clicking it opens Omniscio and navigates straight to that channel and message, Discord-style. The link carries the workspace kind + id, channel id, and message id, so it works across both company and self-serve workspaces.Quote a message — or just the part that matters — highlight any text inside a message — with the mouse, or from the keyboard with Shift+arrow keys — and a small Quote button appears at your selection; click it and just that passage drops into the composer as a markdown blockquote, attributed to whoever wrote it (
> Alex wrote:above the quoted lines), with the cursor on a fresh line below ready for your reply. To quote a whole message instead, use Quote in the message's ⋯ menu (desktop) or the long-press action sheet (phone). Both routes build the identical block from one shared helper, so they never drift apart. Two things it deliberately does NOT do: the attribution is plain text, never an@mention, so quoting someone never notifies them; and the quote is built from the message as you see it, so a mention inside the quoted text comes through as@Bobrather than the raw account id the message actually stores. A quote is ordinary text you can edit before sending — nothing new is stored on the message, so there is no reply-link to break. Quoting is hidden on an attachment-only message (an empty quote is impossible), it lands in the composer for the surface you're in (a thread quote never leaks into the channel box), and it never clears a draft you'd already started. The highlight button is desktop-only — phones deliberately disable text selection in the message list so long-press can open the action sheet — and it has an off switch: Settings → Sessions → "Quote button when you highlight text" (on by default,teamChatQuoteOnSelect) hides the floating button while leaving the ⋯ menu's Quote untouched. Invariants: team-chat-quote-contract.md (the-block-shape-is-fixed…mobile-quotes-from-the-action-sheet); the composer's insert handle isinsert-quote-and-insert-text-are-ref-scopedin team-chat-composer-contract.md.Profile card — click any author's name or avatar to see a small card with their name, photo, role and online status. A Message button (hidden on your own card) opens or creates a DM with that person.
Member directory — a directory of everyone in the workspace (avatar, name, role, presence). In a channel, the member count beside its name opens it when channel invitations are off; a direct message has a Members button instead. Each row has a Message button that opens or creates a DM with them. The header always has an Add members button: for a workspace owner or admin it closes the directory and opens the invite flow (an email invite for the company workspace, a share code for a workspace you created); for everyone else it is grayed out and says why on hover or keyboard focus. The same always-shown, grayed-when-not-allowed rule covers the Invite people link in the New message dialog and Invite new people in Add channel members, so every member list in Team Chat offers a way to add people. Invariant:
member-surfaces-agreein team-chat-desktop-invariants-contract.md.Auto-heal (background safety net) — when you open Team Chat, a one-shot background check silently verifies your member doc exists. If it's missing (you were assigned to the org before Lane A existed, or the backfill was incomplete), the Cloud Function creates it on the spot — no admin action needed, no sign-out/sign-in. It runs once per sign-in — a Team Chat reconnect after a network drop does not repeat it — fire-and-forget, and stays silent on failure. Your app does not repair other people's member docs: an admin's backfill does that, and each person is repaired when they sign in. This softens the hard ordering prerequisite (backfill before gate flip) — a missed user no longer hits a dead end. Invariants:
ensure-chat-member-auto-heals-on-open+no-client-batch-heal-of-org-membersin team-chat-identity-contract.md,opening-heals-membershipin team-chat-desktop-invariants-contract.md.Try again really reconnects — after a failed connection, Try again runs the workspace lookup again (the store counts each press in
orgResolveAttempt), so a connection that has come back leaves Connecting… instead of waiting there forever.A message sent while offline stays "Sending…" until the server has it — on the desktop, a message typed while offline keeps its Sending… label, stays in the saved unsent list, and makes quitting warn you, until the server confirms it. Editing a message that already arrived never counts as an unsent send.
A connection that fails because you are not signed in says so — when Team Chat cannot connect because there is no signed-in Omniscio account, the panel shows Sign in to use Team Chat with a Sign in button that opens Settings → Accounts, and it runs no automatic retry countdown (a retry cannot fix a missing sign-in). A failure that really is temporary still shows Could not connect to Team Chat, counts down to an automatic retry, and offers Try again.
Presence & typing — see who is online / away and who is currently typing (the signal clears itself via a short TTL if a client drops). With Team Chat open in more than one window, only the focused window writes the online/away state; the others refresh the last-seen stamp alone, and "away" is written only once the last focused window loses focus, so the dot no longer flickers between windows.
Profile photos in the DM list — the sidebar Direct Messages list can show each teammate's profile photo next to their name (initials-circle fallback), keeping their online/away presence dot as a small corner badge so photos never hide who's online. A per-user display option, on by default; toggle it via "Show profile photos" in the Team Chat settings window (sidebar header ⋯ menu) or the Direct Messages heading's ⋯ menu. Applies to 1:1 DMs and cross-org connection DMs (channels keep their
#glyph). Setting:teamChatShowDmPhotos. Invariantdm-photos-are-on-by-defaultin team-chat-desktop-invariants-contract.md.A busy workspace no longer redraws the whole channel list on every message — each new message updates only the channel it landed in (so its unread count still moves), and every other channel stays exactly as it was, so the sidebar does far less work while a workspace is busy. Both halves of the channel query now publish together, once, instead of one after the other.
The sidebar is there the moment the app opens — Team Chat keeps each workspace's channel list on your own device, saved for you alone and never carrying message text or previews, so after a restart the channel list paints straight away instead of waiting for the server. It is refreshed in the background, kept for a week, holds at most eight workspaces (about 60 KB each), and is removed on sign-out, when a different person signs in, or when you leave the workspace.
The gallery keeps what you already loaded — switching tabs, flipping between this channel and the whole workspace, or closing and reopening the gallery shows the pages you had already scrolled through at once instead of wiping them and starting over. It quietly checks the first page again in the background: if something new was shared since your last visit, the list refreshes to show it; if that check fails, you keep the saved view with no error. The saved pages belong to the person signed in: signing in as someone else, or switching workspace, starts the gallery fresh, and if you have been removed from the channel the gallery clears and shows an error instead of the media you can no longer see.
Workspace changes made by someone else reach you live — when the owner of a workspace you belong to renames it, hands it to someone else, or deletes it, the switcher's name, the owner-only menu items, and the workspace tile itself update without a restart. If the deleted workspace was the one open, you are moved to your company workspace (or your next workspace).
A deleted or left channel stops receiving alerts here — deleting a channel clears everything this computer had set up for it (its alert routes, webhooks and Session Starter Bot) before "Channel deleted" shows, and still does so when the webhook list cannot be read. Leaving a channel clears only your own alert routes into it and leaves the channel's webhooks and bot alone for the people still in it. When an alert is refused because its channel is gone (deleted elsewhere) or because you are no longer a member (removed elsewhere), the same cleanup runs then, so the alert stops being sent into a dead channel instead of failing every time. A cleanup that fails is logged and never turns a finished delete or leave into an error.
DMs an agent scheduled on this computer show in Scheduled — the Scheduled view now has a Scheduled by an agent section listing the direct messages an agent queued on this machine for the workspace you are in, with the same row and the same confirm-to-cancel flow as your own scheduled messages. Cancelling one from here works whichever session scheduled it, because it is your own computer; an agent's own token can still cancel only what its own session created. The list refreshes itself when a scheduled DM is sent, fails or is cancelled, and shows a retry message if it cannot be read.
Scheduled-DM recovery and delivery bookkeeping run off the app's main thread. Waiting for the local database no longer freezes the window. Status updates are published after the write commits; messages missed while signed out remain failed until you retry them, and signing back in only resumes future messages belonging to that account.
The channel AI summary switch says what it really does — its hint now reads that turning it on sends a sample of recent messages to the AI provider to write the summary, instead of claiming the messages are never read.
A forum feed that connects slowly is never lost — if connecting a forum feed times out after the server already created it, the section reloads its list so the feed appears, rather than showing nothing and refusing a retry as a duplicate.
The once-only forum signing secret is protected — closing the channel editor while a just-created forum feed's signing secret is still on screen asks you to confirm first, because the secret is shown only once and cannot be recovered.
Drafts belong to the conversation you are in — each workspace keeps its own draft, so moving from your company workspace to a workspace you created brings back that workspace's unsent text, and switching back restores what you had there. The company workspace keeps the draft key it has always used, so a draft you already had survives the update, and an older plain-text draft still loads. A draft also remembers the project and channel you picked for posting into a project — come back to it and the chips are still picked, while a pick that no longer resolves is dropped rather than restored as a dangling chip. The draft is also saved when you close the panel or the app, not only while the box stays open.
Names, not account ids, wherever a message is read back to you — editing one of your own messages opens with the people you @mentioned shown as names, and saving turns those names back into the same people (a name two teammates share stays on the person you actually mentioned, and an edit that changes nothing writes nothing). The Saved view and Search results show names too. Replying from an inbox alert suggests the people and channels of the workspace you are replying in, falling back to the company roster for a workspace you have never opened.
An agent lane's daily limit names its reset in your own time — the lane card shows when the daily cap resets as a local day and time (for example Sat 6:00 PM for a reader in Chicago), not as "midnight UTC".
Switching between teammates' agent lanes never shows the last teammate's switch — the lane card only shows a status row that belongs to the teammate you just opened; until that row loads it shows the panel's usual "Checking this lane" state.
Opening a search result is quicker — when results arrive, the top three are read ahead, and any result you rest the pointer on or move to with the keyboard is read ahead too, so opening it usually lands at once. A read-ahead is used once and is forgotten after 30 seconds or on sign-out.
A deleted account's name leaves removal notices too — a notice that names several removed people now records who each one is, so when one of them deletes their account, their name in that notice becomes "a former member". Notices written before this change are not rewritten.
Security and integrity invariants
- Invite codes are server-capped and replace-on-mint. Every code a workspace mints carries a 30-day expiry and a 50-join ceiling, both enforced server-side, so a code is never an unlimited, non-expiring door. In the invite dialog, Generate a new code mints a fresh code and revokes the one it replaces, and Revoke this code ends the code on screen. A code the owner no longer has on screen cannot be revoked: clients are denied any read of the invite-code documents, so the mint callable's return value is the only copy the client ever holds.
- Copy hands over a whole invitation, not a bare code. The invite dialog's Copy button puts a complete, ready-to-send message on the clipboard: the workspace name, what Omniscio is, the download page, the sign-in and Team Chat steps, and the code with where to paste it. Someone who receives just a code has no way to tell which product it belongs to or that Team Chat has to be switched on first. Copy just the code stays beside it for anyone who wants only the string.
- Joining a team lands you somewhere that shows it. The "You joined {workspace}" toast's button reads See your team and opens Settings → Workspaces, which lists the workspace and its members on any install. Switching the active workspace alone changes nothing visible while the workspace switcher and Team Chat are both switched off, so the button never relies on either.
- Non-DM channels are created by the server relay. Creating a channel (anything that is not a DM) goes through the relay, which enforces the workspace plan's channel count and the abuse ceiling before the channel document exists; the client then writes the membership documents and the
memberUidsindex in one batch. A DM id stays a direct client write, because it is deterministic and derived from the participants (dm_<uid>__<uid>), so the relay never mints one. - A typing marker is published only for a public channel the client can prove is public. The marker rides the caller's presence document, which every full workspace member can read, so a DM id or a private channel's id published there would tell the whole company who is privately messaging whom and which private channels are active, live. A channel id the client does not know counts as unproven and publishes nothing; the marker can still be cleared to
null, so a stale public-channel marker does not linger. - Starter-bot configs are keyed by workspace kind, workspace id and channel id. The in-memory cache, the unique index and the message dedup key all use that triple, so a channel id that happens to repeat in another workspace can never read, drive or suppress someone else's bot.
- A Discord import never maps a member by name. A display name is not an identity, so every imported member starts unmapped and an admin confirms each one before the import writes anything.
- Where the starter-bot code lives. The in-memory config cache is its own module (
src/main/services/team-chat/team-chat-starter-bot-cache.ts): it loads lazily, is invalidated by the mutation call sites (IPC handlers and CLI action handlers/routes) and is reset on identity change. Prompt construction, dedup and spend tracking, and spawn orchestration all live inteam-chat-starter-bot-service.ts.
Speed cost
Measured Friday, Nov 6, 2026 on a test copy of the app (a fresh profile on the cloud perf runner, never anyone's live app), with Team Chat off and then on. The on run had Team Chat and its Mission Control link (PM + Team Chat) switched on, Mission Control on in both runs, and a seeded workspace: 12 channels, 6 direct messages and 200 messages of history, every tenth one linking a Mission Control task. Team Chat was opened once and its composer typed in, then both runs settled and idled on the same File Converter screen. One run of each, in the format every feature shares (Speed Cost Format).
| Measure | Off | On | Cost of having it on |
|---|---|---|---|
| Background activity | 12.6 CPU-s/min, 56 calls/min | 17.9 CPU-s/min, 425 calls/min | +5.3 CPU-s/min (about 9% of one core), +369 calls/min |
| Typing delay | median 13 ms, p95 29.6 ms | median 10.8 ms, p95 29.7 ms | none (within noise) |
| Freezes | worst stall 46 ms, 0 long tasks | worst stall 65 ms, 0 long tasks | +19 ms worst stall, no new freezes |
| Memory | 1,336 MB | 2,071 MB | +735 MB |
| Startup | 10,797 ms | 10,720 ms | none (within noise) |
Typing in Team Chat's own composer (on run only): median 26.7 ms, p95 34.6 ms, worst 61.1 ms.
Verdict: Team Chat does not slow typing, startup or responsiveness, and causes no freezes. Its cost is background: about 735 MB more memory after it has been opened, and more than seven times as many calls from the window to the app's back end while you are on another screen. Those two are worth a follow-up that names which calls they are and what holds the memory.
Not measured: real cloud traffic. The test copy replaces Team Chat's cloud database with a local stand-in, so the live listeners, presence writes and their network cost are not in these numbers. Known background work from the code: a presence write every 30 seconds while you are active, a once-a-minute presence check in the sidebar, and a one-second typing-indicator tick while someone is typing. Also not measured: a large real account, video calls, the phone app, and run-to-run spread (one sample per run).
Re-run: the tests/perf/team-chat-speed-cost.spec.ts spec.
Related
This page is split across five files. The parent page you are reading carries the feature's description and everything the first version of Team Chat can do; part 2 covers the second wave of features, more of the everyday surfaces (your profile and manual availability, unread badges, history, jump to date, the quick switcher) and the workspace story; part 3 covers the phone client, notifications and the smaller surfaces around the chat; part 4 covers what the feature does not do and the engineering behind it; and part 5 covers how an owner or admin brings people into a workspace from inside chat.
Forum feed connects a Discourse forum to a channel so every new topic posts there automatically.
Last verified 2026-11-18