Skip to main content
o4 keeps two kinds of lists while it works. The todo list is a checklist of the steps in the current job, which you and the model both edit. Background tasks are commands and agents that keep running while you and the model do other things, such as a dev server or a test watcher.

​
The todo list

For a job with several steps, the model writes a todo list and updates it as it goes. Each item is pending, in_progress, or completed, and at most one item is in progress at a time. When there are open items, the status row above the prompt shows the count, for example 3 todos Alt+L, while the model is idle.

​
Open the todo list

Press Alt+L to open the Todo List pane. Press Alt+L again or Esc to close it. The key is the toggle_todo_list action in [keybindings]. /todo is meant to open the pane too, but in o4 0.2.74 it only refreshes the pane when it’s already open. Use Alt+L. The pane’s footer also shows / filter, but in o4 0.2.74 pressing / opens the command list instead of filtering the todos.

​
Add a todo

Add an item without opening the pane:
Items you add are marked as yours. The model can change their status, but it can’t reword or remove them. An item you delete stays deleted, even if the model sends it again. When you add, advance, or delete an item while the model is working, o4 tells the model right away so it can adjust.

​
How the model uses the list

The model edits the list with the todo_write tool, which runs without asking for approval. Each call sends the whole list. o4 includes the current list in the model’s context on every turn, so the model always sees your changes. When every item is completed, the list clears itself. o4 keeps one todo list per running o4, shared by all of its tabs. The list isn’t saved: it’s gone when you quit, and a resumed session starts with an empty list. It can hold up to 100 items of up to 500 characters each.

​
Background tasks

A background task keeps running while the conversation continues. Background tasks include:
  • shell commands the model starts in the background, such as a dev server,
  • long shell commands and subagents that o4 moved to the background,
  • subagents the model started in the background (see subagents),
  • recurring prompts you start with /loop,
  • campaign workers (see campaigns).

​
Shell commands in the background

The bash tool has a run_in_background option for long-lived commands such as npm run dev or tail -f. The command returns a task ID right away, and the model checks on it later. A foreground shell command that runs too long also moves to the background instead of being killed, so its output isn’t lost. This happens when the command reaches its timeout, or after 2 minutes by default. sleep commands are never moved. To change the 2-minute limit, set auto_background_ms in ~/.o4/settings.json or .o4/settings.json. The same limit applies to foreground subagents.
When a background task finishes, o4 tells the model the next time it can.

​
See and control background work

Run /agents to open the Agent command center. It takes the place of the chat and lists the agents, background tasks, and campaigns in this o4, running or finished. You can open it while the model is working. The line under the title counts the rows: N need input N working N done. Rows are grouped under the tab that started them, and work that belongs to no tab is grouped under Process. A subagent is indented under the agent that started it. Each row shows a status mark, the name, and one status word: The Task details section shows the selected row’s kind (such as sub-agent or background task), status, latest activity, run time, tokens, tool calls, and prompt. It’s a column on the right when the terminal is at least 100 columns wide, and below the list otherwise. The footer also lists g group and n new. In o4 0.2.74, g only cycles the footer label between group: project, group: status, and group: model; the rows don’t regroup. n does nothing when you open the dashboard with /agents. If you open the dashboard with e from the agent and task selector instead, it lists your recent sessions, grouped by project folder, with each session’s running work under it. There, n starts o4 in the selected row’s folder if you trust that folder, and ends the current session first. Alt+B opens the Task Board, a compact table of background tasks and subagents with their kind, status, token use, last tool call, and run time. Press Enter to inspect a row, s to steer it, or k to kill it; k acts at once, without asking. Because k kills on the Task Board, use Up and Down to move between rows there. The footer also lists p review diff, but in o4 0.2.74 p types into the prompt instead. /diff shows the same diff of the session’s changes.

​
How the model manages tasks

The model has these tools for background tasks: task_wait waits up to 300 seconds by default and up to 1800 seconds at most. If it times out, the task keeps running. These approval rules apply in the default approval mode. See permissions for the other modes.

​
Recurring prompts with /loop

/loop sends a prompt or a slash command again and again on a schedule. It runs as a background task in the current session.
You can put the interval first or say it at the end:
Intervals use s, m, h, or d, such as 90s, 30m, 2h, or 1d. Without an interval, the loop runs every 10 minutes. Loops run on a cron-style schedule, so o4 rounds seconds up to whole minutes, and anything over 59 minutes to the nearest whole hour (90m runs every 2 hours, 80m every hour). It tells you when it rounds. The first run starts right away. o4 then shows a Loop Scheduled pane with the task number, cadence, cron expression, and prompt. If the prompt opens a pane of its own, such as /cost, that pane replaces it. A loop is a background task, but in o4 0.2.74 it doesn’t appear in /agents. To stop a loop, press Alt+B to open the Task Board, select the loop’s row (Loop: <prompt>), and press k. A loop also stops when you quit o4.