Agent

Chat, Builder and Auto-Run modes, approvals and the allowlist, checkpoints and revert, plan and progress, and how tool calls appear in the panel.

Three modes

Each session has a mode. Change it from the composer's mode picker or set a Default mode in Settings.

Chat: read-only. Only reading, listing, searching and web fetching. It does not edit files or run commands; it explains the change as code and suggests switching to Builder. MCP tools are not attached in this mode.
Builder: edits files directly. Every turn is checkpointed, and commands not on the allowlist ask for approval before they run.
Auto-Run: edits files and runs commands without asking, all the way to the end. Dangerous commands still ask, even here.

Approvals and the allowlist

In Builder mode these ask first:

Every command (run_command), unless it starts with a prefix on the command allowlist in Settings. The default list is ls, cat, pwd, git status, git diff, git log, npm test, npm run and node --version; add one prefix per line.
Writing or editing files outside the workspace, and deleting files.
Tool calls on MCP servers.

The approval request shows the exact command or path. Commands such as rm -rf /, mkfs, dd if=, a fork bomb, shutdown or reboot are classed as dangerous and never run without approval, not even in Auto-Run.

Checkpoints and revert

In Builder and Auto-Run, every agent turn is a checkpoint. The Revert button beside a turn restores the files that turn changed to their state before it, and tells you how many files went back. You can return to any turn, so a long Auto-Run job you do not like can be taken back in one click. The DiffView tool shows the agent's recent edits before and after.

Plan and progress

For multi-step work the agent writes a short plan with todo_write and updates it as it goes. The Plan strip above the panel shows how far it has got. The built-in Planner preset is set up to produce a numbered plan before any edit.

How tool calls appear

The agent panel is a timeline of text and tool calls in order. Tool calls fold into cards; expand one to see its input and output. File edits carry a per-edit diff, thinking shows as a summary, and each turn shows its token usage. Replies stream in.

The built-in tools are read_file, list_dir, search_by_regex, find_files, write_to_file, edit_file, delete_file, run_command (foreground or background), check_command_status, todo_write, web_fetch and open_preview. open_preview shows a dev server inside the in-app Browser tool. Tools from any connected MCP server are added to this list.

Sessions and agent presets

Sessions are saved per workspace; switch, create or delete them in the panel's session list. History is stored in a provider-neutral format, so switching models mid-session keeps the tool calls and results. Choosing an agent preset in the composer (Planner, Code Reviewer, Test Writer, Docs Writer, 커머스 빌더, or one you made) applies that preset's system prompt and its default model and mode. Presets and rules are covered in Rules and MCP.