Skip to main content
Step Code centers interaction on one input box: type messages, paste images, reference files, and run shell commands. You can also switch approvals, Plan mode, and goals, or send messages while output streams.

​
Input basics

Windows Terminal also supports Ctrl+Enter for newlines. Customize ~/.stepcode/agent/keybindings.json, then run /reload. See Keyboard shortcuts.

​
Pasting images

Use Ctrl+V to paste an image, or Alt+V on Windows / WSL. Supported terminals also let you drag in an image path.
Select a model that accepts images. By default, large images are resized to within 2,000 px to reduce input size; adjust this in [images]. Inline terminal display and sending images to a model are separate capabilities. Disabling display does not prevent image transmission.

​
Slash commands

Type / to open completion; Esc closes it. Invoke Skills with /skill:<name>. See Slash commands for the full list.

​
File references

Type @ to search project files and Tab to complete paths. This works in Git and non-Git directories and on the command line:
Referenced files and files read by tools may enter model context. Check for keys, personal information, and sensitive business data before referencing them.

​
Approval flow

In Ask mode, or when Autopilot encounters an item requiring confirmation, Step Code opens a dialog titled Approve <tool> or Dangerous <tool>. It shows the reason and a summary of the arguments. Approval starts execution; rejection cancels that call and returns the reason to the model. A batch may prompt for each call.

​
Permission modes

Use /permissions to choose, or /permissions cycle and Shift+Tab to cycle. Environment variables can override the initial mode; see Environment variables. Operations identified by dangerous-command rules still require confirmation and are rejected in Read Only. These rules cannot cover every script or dangerous action and are not a sandbox. Non-interactive mode (step -p) rejects calls requiring approval by default. Use --approval-mode to adjust ordinary-call policy; dangerous commands are not exempted. See The step command.

​
Switching modes

​
Plan mode

/plan toggles the mode, /plan <task> starts planning a task, and step --plan starts in Plan mode. The model investigates and writes a plan before implementation. File editing is restricted to read-only investigation, with the plan file as the only writable file. Shell commands still follow normal permissions. Approving the plan exits Plan mode and begins implementation.

​
Direct shell commands

! runs a local command and sends its output to the model; !! runs it without sending the output. Each prefix applies to one command: there is no separate shell mode or user-side background execution. Output appears as a card; long output collapses automatically and can be expanded with Ctrl+O.

​
Goal mode

Goal mode drives multiple turns until the goal is achieved or stopped manually. It suits hours-long tasks. Keep the Step Code process running during the task. The status line shows active time, such as Goal: active · 12m 3s, counting only time when the goal is active. Scheduled tasks use five-field cron expressions. Durable tasks are saved to .stepcode/cron/tasks.json in the project and require project trust. Persisting a task does not let it run without a running scheduler process.

​
During streaming output

  • Enter queues a message for delivery after the current assistant turn’s tools finish.
  • Esc interrupts the turn and restores queued messages to the input box.
  • Alt+↑ retrieves queued messages; use Alt+Q on Windows / WSL.
  • Ctrl+O collapses or expands tool output.
Configure steeringMode and followUpMode in settings. Follow-up delivery has no dedicated shortcut by default.

​
External editor

Ctrl+G edits the current input in an external editor; the action is app.editor.external. Selection order is externalEditor, VISUAL, EDITOR, then Notepad on Windows or nano elsewhere.

​
Next steps