Guide and interrupt agent work
Available in Grida Desktop chat on macOS and Linux. Headless and ACP clients pause, steer and queue through the client operations.
Start from Home
Home and workspace chats share the composer, its + menu, removable Plan and Goal badges, and model and effort picker. Available actions depend on what your installed Desktop and selected connection support. Home has no running turn or waiting queue to control.
Choose the destination workspace on Home, write a request, and send. The workspace opens with a new chat using the model, effort and intents you chose. Drop an image, file or folder anywhere on Home to stage it in the composer; the shared Drop to attach overlay covers the whole Home window. An attached image or file travels with the request. If the connection cannot accept it, Grida keeps the attachment and request in the composer for you to resolve; it does not send the text alone. Reloading before that first send can lose the original selected file. Its card then asks you to attach it again; remove or replace it before sending.
Home also offers reference galleries and slide templates. References become links in the first message. A template is copied into the destination workspace, and the message identifies the copy to adapt. Slides and Images describe the requested outcome; they do not change the agent's permissions or make an unsupported tool available.
Steer the current turn
Send your message while the agent is working. It joins the queue above the composer. When shown, choose Steer on that queued message to have the running turn change direction instead of waiting; the agent uses the correction at its next decision. For example, “Keep the first slide and revise the second one” belongs in the current turn if the agent is still editing the deck.
Steer is not offered for a message queued with Plan, with Ultra effort, as a goal, or with an attachment; those wait for their own turn. If the turn ends before your correction arrives, its text returns to the composer as an unsent draft.
Queue work for later
Send an ordinary follow-up while a turn is running to put it in the queue. For example, “When the deck is ready, describe the changes” can run as its own turn. Each waiting message keeps the options selected when you sent it; a Plan or Ultra choice is shown on the message, but its model is not.
Hover or focus a waiting message to read its full text, including line breaks. Every message keeps its queue icon. An attached image also shows a leading thumbnail when its preview is available. Opening this preview leaves the message in the queue and does not edit it.
Before a waiting message starts, drag it, or open its … menu and choose Move up or Move down, to reorder it. Choose … → Edit to load its text into the main composer. The message stays queued and is marked Editing. Choose Save changes or press Enter to replace its text; its other options stay the same. Cancel or Escape leaves it unchanged. Your unsent draft, attachments and selected modes return when editing ends. If saving fails, your edit stays in the composer to retry. Steering is unavailable while the composer holds an edit. On a supported Desktop runtime, messages with images or files can also be edited: only their text changes, and you can clear the caption while keeping the attachment. The attached content, captured settings and queue position stay the same. Older runtimes offer editing for plain text only. Goals cannot be edited; remove the queued goal and send a new one instead.
If the current turn asks a question, answer it in its card. A separate queued message leaves the question open.
Pause, resume, or move on
While the agent works, the composer's main button (■) pauses it at once and keeps its place. A reply in progress is cut off; tool work already completed is kept. From there:
- Choose ▶ in the empty composer to resume the same turn where it paused. Completed work is kept; a reply that was cut off is written again.
- Or send a new message. The paused turn ends as stopped, and your message runs next, ahead of anything already waiting. Waiting messages follow it.
A turn paused from another client also shows ▶ here. If you paused a goal from its strip, choose Resume goal on that strip instead. Keep an objective across turns
If waiting messages are held after a turn ends, the queue shows Queue paused because you interrupted with Resume. After a full Desktop restart, reopening a chat only loads its saved work; if messages are waiting, choose Resume queue to start them.
Changes already made stay in the workspace. A stopped turn keeps its partial work in view under You stopped. An interrupted model request may still be charged; check your provider's billing record. Inspect and save results
When a turn ends on its own
A turn that fails shows Failed after …. A turn whose outcome Grida could not confirm shows Worked for … · result uncertain; this happens, for example, when the model sends a tool request Grida cannot read, and the turn ends instead of letting the model try again. Work already done stays in the workspace. Review it, then send a new message to continue.
Stop running processes
A command the agent starts can keep running after its turn, such as a development server or a watcher. The composer has no process badge.
In a workspace, the bottom Workspace utilities panel has Terminal and Processes tabs. Processes shows the selected parent chat's agent processes with the same list and Stop controls, including ended entries when none are running. The Terminal tab keeps your separate interactive shell; Processes does not show a live terminal output stream. Switching tabs or closing the panel does not stop the agent's processes.
Type /stop on its own, or choose stop in the / menu, to stop every
process of the chat without sending a message. It does not stop the current
turn; use the composer's main button for that.
Switching chats or reloading the window leaves the processes running, and the
list returns when the chat reopens. Closing the window that shows the chat
stops them, unless another open window also shows that chat. Deleting the
chat or quitting Desktop stops them too. Use /stop in a side chat to stop
its own processes. The workspace's Processes tab follows the selected main chat.
Follow a subagent's work
Click a subagent indicator or named start/finish event in the parent transcript. In a workspace, Subagents opens as a main content tab. Select a child to read its status, recorded activity and conversation. The parent composer remains usable. Closing the tab only hides the view, and another indicator click opens it again. This view offers no creation, assignment or execution controls; give further instructions in the parent chat. Agent process controls live separately in the workspace's Processes tab. Work with subagents
Return to background work
With system notifications turned on, Desktop tells you when a workspace chat finishes, fails or ends with an uncertain result while its window is out of focus. A turn you stopped does not notify, and neither does a question from the agent, so check a backgrounded chat when you expect it to ask something.
Click the notification to reopen the matching chat on its original model. If the model is changing or an attachment is still in the composer, the chat opens after that clears. If the original model is unavailable, Desktop says so before opening the chat. If the workspace chat is still starting, Desktop opens the latest clicked notification once the chat connects.