Composer
Type in Ask anything…. Enter sends; Shift+Enter adds a new line.
| Control | Where | Choices |
|---|---|---|
| + (Add to prompt) | In the composer | Upload files…, Skills, Specialists, Connected Plugs (the plugs on for this chat; GitHub and GitLab insert @github or @gitlab), Lysis investigations (inserted as !name) |
| Mode & Focus (for example Agent · Auto) | Under the composer, left | Mode Agent, Ask, Plan, Debug, Goal (Decisions when a decision model is set up). Focus Auto, Code, Creative, Work |
| Model | Under the composer, right | The model button (Choose model until one is picked), grouped by provider (a Cursor connection splits into Cursor 1st Party and Cursor API); the model’s own options, such as Effort, Fast, Thinking or Max mode; and the optional Decision model. Locked while the chat works |
| Usage ring | Beside the model | How much of the provider’s plan is left; the Profile’s other providers fan out beside it. Click a ring for its limits |
To attach files, use Upload files…, drag them onto the composer, or paste them (a pasted screenshot attaches as an image). Click a thumbnail for a larger preview.
Mode and Focus apply to your next message. Focus guides behavior; it does not change permissions or the model. Delegated tasks run on the chat’s model; when the agent wants another Profile model, a card in the turn lets you keep this chat’s model, use the suggested one, or pick another under Other. See Providers and models.
While the agent reasons, the status line is a ticker: it shows the thought the agent has just finished (or Thinking and a few task words while there is none), one line at a time.
Edit and rewind
Section titled “Edit and rewind”Hover a sent message to see Edit message and Rewind to before this message beside Copy message. They appear only while the chat is idle.
- Rewind chat and files takes the chat back to before that message and
undoes the file edits the agent made since (
Write,Edit,Move,Delete). A file you changed afterwards is kept and named. Commands the agent ran through Bash are not undone; the note under the chat says how many. - Rewind chat only leaves files as they are.
- Edit message rewinds chat and files, then puts the message back in the composer to change and send. Editing the first message opens a new chat with its text instead.
Nothing is deleted: earlier turns stay in the chat’s saved history on its previous branch. There is no way yet to switch back to them.
Drafts, queues, and steering
Section titled “Drafts, queues, and steering”While a response runs and the composer is empty, the send button is Stop response; pressing Esc twice quickly does the same. Once you type, the button becomes Queue message or Steer current response, whichever Settings → General → Conversation (When you send a follow-up during a response) chose as the default. Enter sends it that way; ⌘Enter (on Windows, Shift+Enter) sends it the other way.
- Queue saves the message as the next turn. Queued messages sit above the composer: click one to edit, or choose Steer now or Discard.
- Steer sends text (up to 64 KB) into the current response at its next safe boundary. Attachments and context need Queue. A steer typed while a tool runs reaches the model right after that tool’s result.
- Stop cancels the response and pauses the queue. The composer and your draft stay; a quiet Stopped line offers Resume to continue the turn.
Drafts and queues are saved locally and survive a restart; a reopened queue stays paused until you Resume it.
Work that keeps running shows on the thread rail. The Specialists and
Processes cards show running/total on their title line while something runs, such as
1/3, and just the total otherwise; open one to
list what is running, then what has finished. Processes covers background
commands the agent started, a long command the turn is waiting on, and open
terminals. Each shows by its short name (the agent names a command when it runs
it); a running one’s icon pulses and it has a stop icon. When you stop a
command, the agent is told you stopped it on purpose and not to start it again:
at once if it is working, otherwise with your next message. A stop never starts
a turn by itself. Specialists work the same way: a running one has a stop icon, and one you
stop is told to the agent as stopped on purpose. Background commands keep running while the chat is idle; when
one the agent started ends on its own, the agent is told and the chat shows a
short Terminal notice.
Approvals and questions
Section titled “Approvals and questions”How often Synoec asks is set by the Sandbox level in Settings → Permissions:
| Level | Behavior |
|---|---|
| Strict | Asks before edits, commands, web access and anything else not proven read-only |
| Approve for me · Recommended (default) | Runs web searches, reads, edits and sandboxed commands inside the project. Asks before touching files outside it, publishing, payments and other real risks. On Windows the sandbox limits what commands change, not what they read or send |
| Allow all | Never asks for tool approval, including files outside the project, unsandboxed commands, publishing and payments. Your own deny and ask rules, and consent to control other apps, still apply. Turning it on asks Allow all? first |
Gates sit in the turn, not in place of the composer. When an action needs permission, a Permission needed card shows what it is, with Details, then Always allow, Allow session, Allow once or Deny. The composer stays open: a message you send while a request waits declines it and steers the turn instead. Allow session covers that tool in this chat; Always allow remembers only that action in this project, and an Always allow granted inside the sandbox never covers a run outside it. Clear saved approvals in Settings → Permissions forgets them. The composer is replaced only when nothing could be sent: an interrupted response, a saved queue that needs a decision, a pending message, no model set up, or a context-sharing review.
Two asks come from the sandbox itself. Trust this project for sandboxed
commands? (Trust / Not now) lets commands run without asking and
change that folder but not the rest of your files; trusted projects are listed
under Sandboxed projects in Settings → Permissions. When a sandboxed
command is refused because Git must write to .git, Synoec itself asks, with
the ordinary card, to run that one command outside the sandbox.
A Quick Chat can be given one of your configured projects for its session: the first read there asks (Synoec is asking to read files in alpha), Allow session covers later reads in that project for this chat, and the first edit asks again. Another project, or another chat, asks afresh; the grant ends when the engine restarts or the sandbox level changes.
A turn can also show questions (Your input is needed) as compact choice cards, a Working checklist, a Learning in this chat card (Learn from this?, Accept lesson) and, when there is a decision for you, a Next step card. Its reply buttons send their words as your next message; Not now turns the question down, and Something else puts you in the composer to answer in your own words.
Under an answer, a turn can show generated images (Copy image, Save
as…), the files it wrote with Open and Show in Finder (scripts and
other executables only show in Finder), and N sources from web searches.
Links open in your browser; paths and file: links show as code and never open
from a reply.
A web link on its own line shows as a card with the page’s name and site. Open shows the page in the Workbench Browser (its own tab, or a blank one); once it has loaded there the card shows a picture of it, kept only until Synoec quits and never of a private tab. The card’s ⋯ menu has Copy URL, Open in default browser and, while the page is open beside the chat, Close tab. Outside a chat with a Workbench, Open uses your browser.
A one-line shell block (bash, sh, zsh or shell on macOS; powershell,
pwsh or ps1 on Windows) has Run beside Copy. Run opens a new Terminal
tab in the chat’s project, types the command and presses Return, without
opening the pane; the button then reads Show terminal, which opens the pane
on that tab. Multi-line blocks and lines that start with a $ or > prompt
are copy-only.
For a step only you can do, such as signing in on a website, the assistant
shows the link and waits. Reply Done, or describe the problem, and the same
turn continues. See external steps.