Chat and modes – Docs

Applies to: StablEdge 0.4.4 and later

Chat is how you work in StablEdge most of the time. Pick a mode to have the Agent answer a question, propose a plan first, or go ahead and act; you can also attach files, call Skills, and split subagents off to work in parallel.

Choosing a mode

Switch in the mode selector above the input box, which shows the name of the current mode. A switch takes effect when you send the next message.

Mode	What it is good for
Agent Lite	Acts directly; complex research tasks escalate to several Agents on their own. The first choice for everyday work.
Agent Complex	Long complex tasks: break the work into a todo list, carry out and check each item, then deliver a summary.
Chat	Read-only questions: search and read files, without changing them or running commands.
Plan	Produces only a plan and todos, and waits for your confirmation before acting.
Task	Build an automation task inside the chat and save it to Task Center in one click; see Building a task from a chat.
Focus	Watch a window or screen on this computer and handle changes the way you ask; see Focus mode.

If you switch away from Agent Complex while a task is running, the task is abandoned when you send the next message. Messages and files already produced are not deleted.

Using Plan mode

Choose Plan, describe your goal and send it.
The Agent returns a plan and lists its assumptions; anything it needs from you is marked as open, and the work continues only once you have answered everything.
Select Execute plan to start, convert it to todos, or edit the plan and generate it again.

The input box

The placeholder reads "Ask the AI anything. # for Skills, @ to mention colleagues".

Above the box, from left to right, sit safety confirmation, the mode selector (showing the mode name), the project and Abilities. Inside the box, the bottom left has attachments, model choice and Deep thinking.

Sending and line breaks: Enter sends and Shift+Enter starts a new line. For longer text, select Expand input box.
Skills: type # and pick a Skill to insert. You can also use Abilities above the box to pick a Skill or enable a plugin for this chat.
@ mentions: type @ to pick a colleague, who is notified when this turn ends.
Attachments: select Add attachment and then Add file or Add folder, or drag a file into the box. Past conversations from the sidebar, chat tabs at the top, and page and file tabs can be dragged in as references too.
Model: switch this chat's model in the model selector inside the box, after the attachment button.
Deep thinking: sits after model choice. Turned on, the Agent reasons at the effort set for the model; turned off, it answers straight away.
Safety confirmation: the leftmost control above the box lets you choose Require safety confirmation or No safety confirmation for this chat. With it off, high-risk actions run directly, so use it with care.

Drafts are saved automatically.

While a reply is being written

Keep sending messages: you can type and send before the reply ends; messages join the queue, which shows "N queued, sent in order when this turn ends". You can edit or delete them, drag to reorder, or choose Send now and Insert this message. If the previous turn was stopped or failed the queue pauses, and you can continue it by hand.
Stop: select Stop to end this turn.
Withdraw: click your own message bubble. While there is no answer yet the turn is withdrawn and the text goes back into the input box; once there is an answer, the message opens for editing and you can send it again.

Confirmation and question cards

With safety confirmation on, a confirmation card appears before the Agent changes a file, runs a command, or uses or installs a Skill. You can:

"Allow once": lets this one action through;
"Allow {category} this turn": stops asking for the same kind of action, such as file changes or commands, for the rest of the turn;
"Allow all this turn": stops asking about anything for the rest of the turn, though installing Skills and calling plugin workers are still confirmed individually;
"Refuse": does not run the action.

A confirmation card expires if no one answers it for a while; the wait can be adjusted under Settings → Agent.

When the Agent hits a question only you can settle, a card headed "Your answer is needed" appears: pick an option or write your view and submit. If the question expires you can still send your answer by selecting Send as new message.

When a website needs a sign-in, an account and password card appears. The password is stored encrypted in the vault on this machine only and is never sent to the model.
When an API key is needed, a secret entry card appears. The key is kept in the system keychain only; see Secret environment variables.

Reviewing and undoing changes

Above the input box a summary lists the files this chat changed and the lines added and removed. Keeping or undoing here never commits anything to your Git repository.

Select Show changes to look at the diff. Binary files and files over 2 MB show no diff.
Keep or undo one file at a time, or select Keep all or Undo all.

You cannot undo or keep while an answer is being written; wait until the turn ends.

Reply actions and branching

Hovering over an Agent reply shows icon-only buttons below it: copy, quote and branch from here. Branching copies the conversation up to that turn into a new chat so you can take a different direction.
Select text inside a reply to copy it into the input box, or to ask about it in the box beside.
A chat tab's context menu also offers Branch conversation and Branch in a separate worktree. While you are in a worktree the top of the chat shows "In a separate worktree · branch …"; when you are done select End worktree, and the changes are committed to that branch for you or the Agent to merge.

Subagents

For larger tasks the Agent can send out subagents to work in parallel, and the chat shows a card and progress for each subtask.

Subagents run in the background by default and wake the chat when their result arrives. Select View subagent run to follow one in its own tab, and add requirements while it works.
To stop a single subtask, select Stop; the main chat carries on.
A subagent that may change files works in a separate worktree; once the main Agent has reviewed and merged, apply everything to the workspace in one step above the input box, or undo it as a batch.

Context and compaction

The percentage beside the input box shows how much context is in use. When it is nearly full, StablEdge condenses the context in the background: earlier messages are summarized and the most recent turns are kept as they are, so you can carry on. You can also select Compact context yourself.

If the main model keeps failing, runs out of quota or goes offline, the turn continues on the fallback model chain from Settings, with a notice in the chat, and the next turn returns to the model you chose.

Focus mode

Focus mode has the Agent watch a display, app window, terminal or log file on this computer. When the screen changes, a command fails or a log line goes wrong, the Agent handles one round and gives you a summary, and stays quiet when nothing needs doing.

Choose Focus in the mode selector above the input box.
Select Add or change view and pick the display, window, terminal or log file to watch.
If you need to, open Masking and cropping to box the area that matters or cover parts that keep moving, such as a clock.
Write what you want done in the input box and select Start focus.
When you are done, select End focus. You cannot switch modes while focusing.

With Allow execution turned on, the Agent can fix problems in the terminal itself, and high-risk commands need your confirmation every time.
While watching a display you can turn on Plain wallpaper during focus to cut down distractions.
With Remote delivery turned on, focus results are pushed to the Remote Control channels you have connected; see Managing focus from your phone.