Terminal keys

The input line is a small text editor, and it does not behave like a shell prompt.

Three habits of the terminal are deliberately gone: Ctrl+C copies the selection instead of killing the program, Ctrl+D deletes a line instead of ending the session, and Ctrl+Z takes an edit back instead of suspending the process. Ctrl+Q is the way out. The editing keys are the same as in idet, so the two do not have to be held apart.

Four of the keys below belong to a command rather than to the editor — Alt+H, Alt+T, Alt+G and Ctrl+Q — and those are the ones a Shortcuts section in settings.cfg reassigns, described under configuration. The help panel shows each command beside the key in force for it, so a reassignment reads back where it was made.

Sending and stopping

A turn stopped with Esc keeps what the model had already said, up to the last complete word.

The history above the input

Tabs, panels and the goal

The log scrolls with the mouse, and a click unfolds things that are folded: the diff under an edit, a block of the model's thinking. While the view is scrolled up, the bottom row says how far back it is and takes a click to come back. The first and last rows of the log name the user message above and below and scroll to it.

Because a terminal hands its selection to the running program as a property of the whole window, an application that wants the mouse in one place gives up the terminal's own selection everywhere. Revolve needs it in the input field, so it draws the selection over the log itself. Shift+drag is what your terminal offers when you want its selection back for a moment.

Recalling what you typed

Recall is a mode rather than a position. Once Up has produced an earlier message, Up and Down step between messages directly; any edit or cursor move ends that, and the next Up starts again from the newest. With text already typed, only the entries containing it are walked, the way fish does it.

What is recalled is what you actually submitted, kept separately from the conversation, so a heartbeat note or a message from another agent never shows up as something you typed. The grey autosuggestion at the end of the line draws on this session first and then on earlier sessions in the same directory; it never changes the input on its own. Word suggestions come from the conversation itself, so a file name the agent just printed can be completed rather than retyped.

Editing

What the two status rows say

The first row is what changes from keystroke to keystroke. The second, nearest the input, is the identity: the model and the standing goal from the left, the provider's rate windows and then the context reading pushed against the right edge, so a number you look for often is always in the same place. The context reading turns yellow at 150k tokens and red at 200k, which is warning enough before a compaction happens.

What a turn is currently doing — the spinner, the state word, the elapsed time, a retry countdown — sits at the bottom of the log rather than in the status bar, under the newest output where you are already looking.

How answers are drawn

Replies stream in as they are generated and are rendered as markdown: headings, emphasis, lists and quotes are styled, and fenced code blocks are syntax-highlighted. The highlighter reads its language files from ~/.config/idet/syntax and brings Rust, TOML, JSON, and shell along built in.

A turn's thinking is one entry rather than one per line: while it streams, only its newest line is drawn; once the answer starts, it folds to a single row saying how many lines it was, and a click opens it. A long reasoning pass therefore costs one row instead of pushing the conversation off the screen.