Skip to main content

Live: the agent chat

The Agent panel (Mod+1), also called Live, is where you tell an agent what to do and watch it work. It reads like the terminal beside it: the conversation fills the panel, and the prompt is a command line at the bottom that grows as you write.

Writing a prompt​

  1. Put the cursor on the prompt line: click it, or press Mod+L.
  2. Write what the agent should do. Shift+Enter makes a new line.
  3. Press Enter (or Mod+Enter), or click Run.

Under the prompt line:

ControlWhat it does
Agent pickerChooses the agent that takes the prompt: Claude Code, Codex, Gemini or one you added (see Agents).
waitsHow long the agent waits for your answer when it asks to connect somewhere the policy puts to you: 1, 5, 15 (the default) or 30 minutes, or 1, 2, 4 or 8 hours. Unanswered, the connection is refused and the agent goes on without it.
Measure barWhile a run goes on: what it takes of the computer, against the task's limits. See below.

Places you picked in the Preview to ask about show as chips above the prompt line; see Files and Preview.

One agent works in a task at a time. To run agents side by side, make more tasks, on the same folder or others. You can run the agent again before you commit: changes pile up in the copy until you decide.

What Live shows​

As they come:

  • Your prompt, in a block of its own under You, and the agent's answer under its name. Each message has a copy button (it shows when the pointer is over the message) that copies it as it was written, Markdown and all.
  • The agent's answer, laid out as Markdown, with code blocks coloured as in an editor.
  • Its thinking, in a quieter style, when the agent shares it.
  • Tool calls, one line each: · while running, ✓ done, ✗ failed, and after the arrow the files the call changed. Every call that changes something is a step you can look at and go back to (see Timeline).
  • What each call printed, such as a command's output, under its line:
    • it streams while the command runs;
    • it is open while the call runs or when it failed, and closed once it went well (click the line to open or close it);
    • terminal colours and cursor moves are taken out, and the last 64 KB of each call's output are kept.
  • Notes: the agent's plan, permissions it was given, a server Kule stopped, pictures an agent could not take.
  • Question cards when the agent's program wants to connect somewhere the policy does not let through on its own (see below).
  • The end of the run: Done, Stopped: … with the reason, or The run failed: …, with the number of operations the sandbox refused. Where Kule knows why a refusal keeps a tool from working, it says so under that line. For example, a program that wants the screen (LibreOffice, even with --headless) cannot run in a task, and macOS may report that it quit unexpectedly. The agent is told the same at its next prompt, in Kule's words, once.

The daemon keeps a task's latest events, so a window opened later shows what happened meanwhile.

Copy on select

Text you select with the mouse in Live goes to the clipboard at once, and a small Copied shows by the pointer. Turn it off in Settings → Appearance → Copying.

Questions​

When the agent's program connects to a host the policy does not let through on its own, the connection waits and a card shows in Live. The task also moves under Needs you in the Tasks panel, and, while the window is not in front, the system's notifications tell you.

The card names the host and port, the rule that made it wait, and a countdown to when it is refused.

ButtonKeysWhat it does
ApproveMod+Shift+YLets this one connection through.
Approve and remembernone by defaultLets this host and port through on this folder from now on, without asking. Settings → Rules lists what you told Kule to remember, and forgets a rule.
DenyMod+Shift+NRefuses it. The agent goes on without it.

How long a question waits is the waits setting under the prompt (15 minutes unless you change it, from 1 minute to 8 hours; never off). See Network and questions.

The measure bar​

While a run goes on, the bar under the prompt shows what it takes of the computer, updated every two seconds, against the task's limits. For example:

CPU 140% / 200% · 1.2 GB / 4 GB · disk +350 MB · 7 processes
  • It turns amber near a limit.
  • It adds slowed while the processor limit holds the run back.
  • Click it to change the task's limits.

When a limit stops a run, Live says which one and what it measured, for example "Stopped: the run's processes held 4.2 GB, over its memory limit of 4 GB". If that task is not the one on screen, a notice says so.

Stop​

While a run goes on, Stop (Mod+Shift+K) takes the place of Run. It cancels the agent's turn, as you would in the agent itself. The button reads Stopping… until the run ends; an agent that does not end its turn within ten seconds is killed. Whatever it changed so far stays in the copy for you to review.

After the run​

When Live says the run is done, open Review (Mod+2) to see what changed and commit or roll back. See Review.

From a terminal​

kule run gives an agent a prompt from a terminal, held to the same budgets and limits. kule steps lists what each tool call changed. A terminal has nobody to ask, so a connection the policy puts to a person is refused at once there.