Composing Messages and Attachments

​

Sending messages

• Enter sends the message; Shift + Enter or ⌘ + Enter (Ctrl + Enter on Windows) inserts a line break
• The Enter that confirms a text conversion in IME-based input (such as Japanese) never sends
• You can switch the send key to ⌘ / Ctrl + Enter under “Sending messages” in Personalize settings; in either mode, the Enter combination that is not the send key inserts a line break. The keycap chip next to the send button shows the active key. On phones and tablets, Return inserts a line break — use the send button to send
• Messages sent while the AI is still working wait their turn: the text leaves the input box and sits above it as a pale bubble until the current reply ends. Up to 5 messages can wait — a sixth stays in the box with the note “Up to 5 messages can wait — cancel one or wait for a turn to finish” — and they go out one per turn, oldest first, so you answer each reply before the next message is sent. The ✕ on a bubble (“Cancel queued message”) puts that message back in the box to edit or drop; “Interrupt” returns the whole stack. A message still waiting when you reload or close the tab is not sent from the closing page; it is back in the input box when you reopen the chat. When something other than the running reply is holding a message — a card waiting for your answer, an attachment still uploading or failed — a one-line notice in the box says so. The AI usually opens the reply by noting which of your messages it is answering

​

Reply suggestions

After the AI finishes responding, a predicted next reply may appear as gray ghost text in the empty composer.

• Press Tab to accept it; it is inserted into the composer for you to edit or send
• Start typing your own text and it disappears
• The suggestion is kept with the reply, so it is still waiting for you after a reload — and it is prepared even when the AI finished the reply while you were on another screen

​

Reference pickers

Typing certain trigger characters in the composer opens a picker that lets you search agents, connectors, knowledge, and skills by name and embed them in your message as references. Flows in the # picker are organized by their jinba-flow folder structure — drill into folders, or just type to search across everything (results show the folder path). In a chat bound to an agent whose Jinba Flow connection is narrowed to some of its flows, the # picker and the ”+” menu offer only those flows, and the counts reflect the selection (Attach resources). The [[ picker is a navigable folder tree too: “Add entire folder” attaches the folder’s whole contents, subfolders included. In a chat bound to an agent, the tree is that agent’s knowledge as it is organized in the agent’s configuration panel — its own folders plus the workspace folders its attached items live in. For a workspace folder only partly attached to the agent, “Add entire folder” ticks the visible items one by one instead of attaching the whole folder. Searches can include spaces, so multi-word names like “Weekly Report” match as you type. Letter case, full-width and half-width characters, and hiragana and katakana all match either way (ＡＩ finds “AI”, ｶﾀｶﾅ finds カタカナ, せーるす finds セールス) — the same rule applies to the other list searches in the app: the chat search, the agent, skill, knowledge, connector, and member lists, and the workspace switcher (searching inside an open document, such as a spreadsheet preview, is not covered). The picker closes on a line break, on two consecutive spaces, or when a spaced query no longer matches anything — and Esc closes it too. Selected references appear as chips and can be removed with the ✕.

​

The ”+” menu

The ”+” button to the left of the composer offers the same references, plus attachments. Its rows — and the agent configuration cards — are ordered to match the sidebar, and each picker keeps its search box and an “Add / manage” link at the top.

• Select an agent — bind the chat to an agent
• Flow — pick a jinba-flow connection, an individual flow, or a custom MCP server
• Knowledge — attach a registered file, or a whole folder as one reference
• Skills — pick a skill for the AI to use
• Prompt templates — start from a saved prompt
• Add files or photos / Attach folder — attach files, or a folder’s contents in one go

On a wide window, choosing Flow, Knowledge, or Skills opens that picker as a panel beside the menu. On a narrow window (about 972 px and below — phones, tablets, and a narrow desktop pane) the menu is a single popup as wide as the message box, and the picker opens in place of the list with a back row along the top that reads as a path — “All / Knowledge”, then the folder you have drilled into. Click the back row, press Esc, or press ← in an empty search box to go up one level; Esc closes the menu only from its first screen (clicking outside closes it from any level). On those narrow windows a flow’s or skill’s description is shown as a second line under its name instead of in a hover card. On a wide window, hovering over a flow — in the ”+” menu’s Flow list or in the # picker — shows what the flow will ask for before you pick it: an “Inputs” section lists each input by name with its description, marks required ones with “(required),” and names the accepted file types for file inputs. Up to 8 inputs are listed, with a “+n more” line beyond that; a flow that needs nothing reads “This flow requires no inputs,” and one that carries no description of its own reads “No description set.”

​

File attachments

Attach documents or images you want analyzed.

• Drag and drop files or folders onto the composer, or choose from the ”+” menu
• Attaching a folder (via “Attach folder” or drag and drop) attaches the files inside it individually — subfolders are included, but the folder structure itself is not preserved, and OS bookkeeping files such as .DS_Store are skipped automatically
• There are limits on the number of attachments per message and on file size (an error message will let you know if you exceed them)
• You cannot send until the upload finishes. Please wait for it to complete
• A failed attachment states the reason on its chip — an unsupported file type, over the size limit, an empty file, a server error, a network error, a timeout, or the organization’s storage limit — and the retry button next to the ✕ re-runs that upload in place, so you never have to pick the file again
• An upload that takes longer than five minutes is given up and lands in that same failed state, so the send button is never left disabled with nothing you can do about it

​

Pasting large text

Pasting a long text (roughly 2,000 characters or 40 lines and up) condenses it into a compact “Pasted text” card above the input instead of flooding the editor.

• Nothing is uploaded — the card’s content is merged back into the message when you send
• Remove a card with the ✕; cards are kept with your draft if you navigate away
• Extremely large pastes (1,000,000 characters and up) are attached as a pasted-text.txt file instead
• A part of a preview you pick with “Select & ask” arrives as the same kind of card, labelled “Selected: …” (AI Activity and the Workspace Panel). Up to 20 cards can wait above the input at once

​

Model and options

The bottom of the composer shows the AI model used for this conversation. When you select an agent, the badge follows that agent’s configured model rather than the workspace default; an ad-hoc chat shows the workspace default. Hover over the badge to see which it is. The first time the badge names a different model than you last saw on this device for that agent (or, in an ad-hoc chat, for the workspace), a one-time callout opens on it — “Your AI model moved from <previous> to <model>.” — with an “Open settings” link to the workspace’s default-model card in an ad-hoc chat; it closes by itself after 15 seconds, and a model you changed yourself in the settings is not announced. (The model that actually runs resolves in the order chat override > agent > workspace default.) Beside it, the “Response quality” chip switches between “Standard quality” and “High precision” — the latter re-works the answer before showing it, which is slower and uses your allowance faster; hovering a row in either menu explains the trade-off, and the choice stays on for the following sends until you switch it back. On a narrow composer the agent and quality chips fold into one “Options” menu.

​

Selecting an agent

• Choosing an agent with the agent selector in the composer (or @) applies that agent’s skills, connectors, knowledge, and memory to the conversation. The selector’s list has a “Search agents” box and an “Add / manage agents” row that opens the agent catalog
• Choosing “No agent” gives you an ad-hoc chat
• You can also simply ask to talk to a specific agent, with or without an agent selected — the AI shows an agent hand-off card that starts a new chat with that agent (In-Chat Forms and Suggestions)
• You cannot switch agents after a chat is created. To use a different agent, start a new chat